CN106294489B - Content recommendation method, device and system - Google Patents

Abstract

The application discloses a content recommendation method, device and system. One embodiment of the method comprises: acquiring user attribute information and/or environment attribute information; and synthesizing the recommended content based on the user attribute information and/or the environment attribute information. The implementation can provide more information in the recommended content and improve the utilization rate of the display position of the recommended content. And more targeted personalized content can be provided, so that the conversion rate of the recommended content is improved.

Description

Content recommendation method, device and system

technical field

The present application relates to the field of computer technologies, in particular to the field of terminal technologies, and in particular, to a content recommendation method, device, and system.

Background technique

Existing content recommendation systems (such as advertisement recommendation systems, information recommendation systems, etc.) mainly recommend relevant content based on the analysis of detected users' interests, preferences, and areas of interest. This kind of content recommendation system mainly includes three modules: user detection and feature analysis module, recommendation module and display module.

In the user detection and feature analysis module, after the image is collected by the camera, pedestrian detection, face detection and other methods are used to extract the target user in the image, and then the feature analysis of the target user is performed. The most commonly used features in the prior art include the user's gender, age, clothing color and style, and shallow features such as the user's expression.

The recommendation decision of the recommendation module mainly includes two ways. One way is to make recommendations according to preset rules. For example, when the gender of the user is detected as female, the system will make recommendations based on the preset female attention content. Specifically, it can recommend products such as cosmetics and fashion to women, or Recommend health and beauty information to women. Another way is to recommend through learning. The user's features and the features of the content to be recommended can be vectorized, the two features are associated offline through machine learning, and the user's features are mapped to the features of the content to be recommended through the trained model online, and the features of the content to be recommended are mapped. Match in the content database, so that the content with high matching degree is used as the recommendation result. Among them, preset contents are stored in the content database, the form and the amount of information contained in the preset contents are small, and the form is fixed.

In the prior art, the display module generally has only one display screen and is designed to serve only one group of users. That is, only content recommended for a group of users is displayed within a period of time. The group of users may contain multiple individuals, and by extracting common features among the multiple individuals, such as extracting common age groups, social relationships, etc., content is recommended based on the common features.

There are the following problems in the above-mentioned prior art:

In the user detection and feature analysis module, only the shallow features of users are analyzed and extracted. In practical applications, it is difficult to judge whether products or information are suitable to be recommended to users through these shallow features. For example, when the user is detected as a young woman, the system judges that the user may be interested in cosmetics, but cannot judge what should be recommended to the user. Brand or what type of cosmetic.

In the recommendation module, preset content is stored in the content database, and the recommended preset content has the following defects: First, the recommendation module may decide multiple preset contents to be recommended, and the correlation between multiple preset contents may be poor , by displaying multiple preset contents to be recommended, users will feel a sense of disconnection when obtaining information; second, if the matching degree between user characteristics and the characteristics of the contents to be recommended is very low, the recommendation effect is poor; third, now The content forms and information recommended in the existing technology are fixed, and cannot provide personalized recommended content to meet the needs of different users. For example, when recommending a watch, it is accompanied by different characters and music. Not the same, the recommended effect is completely different.

In the display module, the prior art only serves a group of users, which is difficult to meet the personalized needs of multiple users, and the amount of recommended content is small. When the number of users is too large or the features are too complex, the system may be completely unable to decide what to recommend.

SUMMARY OF THE INVENTION

In view of the above-mentioned defects in the prior art, it is desirable to provide a personalized content recommendation method. Further, it is also expected that the recommended content may contain richer information for multiple groups of users with different characteristics. In view of this, the present application provides a content recommendation method, device and system.

In a first aspect, the present application provides a content recommendation method. The method includes: acquiring user attribute information and/or environmental attribute information; and synthesizing recommended content based on the user attribute information and/or the environmental attribute information.

In some embodiments, synthesizing the recommended content includes: determining candidate content elements based on user attribute information and/or environment attribute information, wherein the candidate content elements include candidate object elements and candidate scene elements; and synthesizing the recommendation according to the candidate object elements and the candidate scene elements content.

In a second aspect, the present application provides a content recommendation apparatus. The apparatus includes: an acquisition unit configured to acquire user attribute information and/or environmental attribute information; and a synthesis unit configured to synthesize recommended content based on the user attribute information and/or the environmental attribute information.

In some embodiments, the synthesis unit includes a determination subunit configured to determine candidate content elements based on user attribute information and/or environmental attribute information, wherein the candidate content elements include candidate object elements and candidate scene elements; and a synthesis subunit, configured It is used to synthesize recommended content based on candidate object elements and candidate scene elements.

In a third aspect, the present application provides a content recommendation system. The system includes a processor and a display device; wherein the display device is configured to display recommended content; the processor includes the content recommendation apparatus according to the second aspect of the present application.

The content recommendation method, device and system provided by the present application synthesize or generate recommended content based on user attributes and environmental attributes. It can automatically recommend personalized content, and the recommended content contains more information, which improves the pertinence of the content recommendation system. At the same time, it can recommend content that meets the user's needs and interests that cannot be directly judged from the user's shallow features, which improves the utilization rate of the content recommendation system.

Description of drawings

Other features, objects and advantages of the present application will become more apparent upon reading the detailed description of non-limiting embodiments made with reference to the following drawings:

FIG. 1 shows an exemplary flowchart of a content recommendation method according to an embodiment of the present application;

FIG. 2 shows a schematic diagram of the effect of individual analysis according to an embodiment of the present application;

FIG. 3 shows an exemplary flowchart of determining candidate content elements according to an embodiment of the present application;

FIG. 4 shows an exemplary flowchart of a method for synthesizing recommended content according to an embodiment of the present application;

Fig. 5a shows a schematic diagram of an effect of displaying recommended content by means of space division presentation;

Fig. 5b is a schematic diagram showing another effect of displaying recommended content by means of space division presentation;

FIG. 6 shows a schematic diagram of the principle of displaying recommended content in a time-division presentation manner;

FIG. 7 shows a schematic structural diagram of a content recommendation apparatus according to an embodiment of the present application; and

FIG. 8 shows a schematic structural diagram of a content recommendation system according to an embodiment of the present application.

Detailed ways

The present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the related invention, but not to limit the invention. In addition, it should be noted that, for the convenience of description, only the parts related to the related invention are shown in the drawings.

It should be noted that the embodiments in the present application and the features of the embodiments may be combined with each other in the case of no conflict. The present application will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

In the following description, numerous specific details are set forth in order to provide a thorough description of embodiments of the present invention. However, it should be understood by those skilled in the art that the embodiments of the present application may be practiced without these specific details.

Please refer to FIG. 1 , which shows an exemplary flowchart of a content recommendation method according to an embodiment of the present application. For ease of understanding, in this embodiment, an example is given in conjunction with a device for displaying recommended content. Those skilled in the art can understand that the device for displaying the recommended content can be an electronic large screen, such as an electronic screen disposed in an underground passage or a shopping mall hall, or a mobile electronic device with display functions such as a mobile phone and a tablet computer.

As shown in FIG. 1, in step 101, user attribute information and/or environment attribute information are acquired.

In this embodiment, the user attribute information can be extracted based on the image captured by the camera. The camera may be installed on the device for displaying the recommended content, or may be installed in one or more locations near the device for displaying the recommended content. In some implementations, images collected by multiple cameras may be acquired, and user attribute information of all users within the display range may be analyzed.

User attribute information may include user individual attribute information and group attribute information. Wherein, the individual attribute information may be information obtained by analyzing the individual characteristics of each user, and the group attribute information may be information obtained based on the relationship between multiple users.

In some implementations, the individual attribute information may include the user's individual characteristic information, such as the user's appearance characteristic information, that is, information that can be directly derived from the user's surface characteristics, such as the user's gender, age, race, clothing style, makeup style , health status, posture and other information. Among them, gender, age, race, and makeup style can be obtained by classifying the features of the face region through a classifier, and health status can be obtained through a classifier or retrieval from the features of the face region and posture. The individual feature information may also include the user's emotional state information, such as happiness, sadness, anger, etc., which can also be obtained by analyzing the user's facial expressions and body movements by using a classifier.

Further, the individual attribute information may also include the user's character and purchasing power. Personality can include at least one of the following: sense of responsibility, emotional stability, extroversion, openness to new things, agreeableness, popularity, confidence, and loneliness. These personality indicators can be quantified into multiple levels, each corresponding to a different degree of conformity. For example, there are 7 grades from -3 to +3, where -3 can represent the least fit, and +3 can represent the most fit. Taking the recommended product as an example, when the user's openness to new things is quantified as +3, some novel products can be recommended to the user, such as a different style or color of clothing worn by the user, or travel information is recommended; On the contrary, when the user's openness index to new things is quantified as -3, products that are consistent with the user's current state, such as clothes and accessories of the same style, can be recommended to the user.

The user's personality can be obtained by analyzing the user's characteristics through a regressor. In some implementations, the personality can be captured by machine learning methods. For example, a training set can be established based on the existing data, and a character model can be trained by using a machine learning method to obtain a mapping relationship between user characteristics and characters. Optionally, the personality model can be trained by using the artificially quantified personality index as training data. For example, you can obtain images of multiple users, extract user characteristics, and then quantify the degree of extroversion of each user into seven levels of -3, -2, -1, 0, 1, 2, and 3 based on psychological analysis. -3 means the least extroversion, +3 means the most extroversion. For example, a user wearing brightly colored clothing can quantify that user's degree of extroversion as +2 or +3. The user's characteristics and quantified extraversion degree are used as training data, and machine learning methods such as support vector machines (SVM) and random forests are used to train the extraversion degree model, and the mapping relationship between user characteristics and extraversion degree is obtained. When judging the degree of extroversion of a user, the extraversion degree model can be used for analysis. In a further implementation, training can be performed based on multiple characteristics and multiple personality indicators of the user, and the obtained model is a comprehensive personality model, and the judgment results of multiple personality indicators of the user can be directly obtained through the integrated personality model.

The purchasing power information can be obtained from the price information of the clothes, shoes, and accessories worn by the user. First, the characteristics of the user's clothes, shoes, and accessories can be extracted, and search and matching are performed in the database to obtain brand information and/or price information of the clothes, shoes, and accessories. Optionally, the price level of the price information in the price points of all similar commodities can be calculated, so as to determine the purchasing power of the user. For example, the characteristics of the watch worn by the user can be extracted, and the brand information and price information of the watch can be searched in the commodity database. Further, it is also possible to query the price range information of watches of the same brand, and sort the price information or price range information among all watch products (for example, the ranking percentage, etc.). Alternatively or additionally, purchasing power may be quantified, eg, into multiple levels. Specifically, if the price of the watch ranks in the top 10% of all types of watches, the purchasing power of the user can be determined as the highest level.

The group attribute information of the user may include the social relationship information of the user, including family relationship, lover relationship, friend relationship, and the like. The social relationship information of the users can be determined from the clothes worn by the multiple users, for example, it can be determined that the multiple users are in a couple relationship or a family relationship through a couple outfit or a parent-child outfit.

In some optional implementations, the user attribute information can be obtained by the following methods: collecting images of the area where the recommended content is displayed, determining the users who are the service objects from the images, and performing individual analysis and group analysis on the users who are the service objects. analyze.

In the above implementation manner, an image of the area where the recommended content display position is located may be collected by an image collection device (eg, a camera, a camera component on a mobile terminal). Optionally, before capturing an image, a sensor may be used to detect whether there is a user and the user's position in the area where the display location is located, and a computer system may control the camera to rotate and focus to capture the detected image of the user.

Optionally, determining the user as the service object from the image may include: detecting a pedestrian in the image and the pedestrian's focus position of sight; judging whether the focus position of the pedestrian's sight is at the display position of the recommended content; if so, Then determine the pedestrian as the user as the service object. After the images are collected, pedestrian detection can be performed based on skin color features or human body shape features, or machine learning methods such as random forests and hidden Markov models can be used for pedestrian detection to extract the human body in the image. Then, based on the color feature (such as black) and shape feature (approximately circle) of the pupil, the pupil position of the human body can be detected by methods such as edge extraction, Hough transform, etc., so as to determine the position parameter of the pedestrian's sight focus, and based on the position parameter Determine whether the focus position of the pedestrian's sight in the image is at the display position of the recommended content. If yes, the detected pedestrian can be determined as the user who is the service object. It should be noted that the display position of the recommended content may include an area, and when the focus of the pedestrian's sight is located in this area, the pedestrian can be considered as a user serving as a service object.

In some optional implementations, the individual analysis of the user may be performed by: dividing each user in the image into multiple sub-images according to body parts, and using a classifier and/or a regressor to analyze the sub-images to Get user attribute information. Specifically, the user detected in the image can be segmented according to different parts of the human body. For example, a human body image can be segmented into a face image, a limb image, and a body image. Each sub-image can then be analyzed using a classifier and/or a regressor. For example, a makeup style classifier can be used to classify facial images, and a clothing style classifier can be used to classify limb images and body images, so as to obtain various attribute information of the user.

Further reference is made to FIG. 2 , which shows a schematic diagram of the effect of individual analysis according to an embodiment of the present application. As shown in Fig. 2, the extracted user image can be segmented into hair image, face image, left arm image, bag image, left leg image, skirt image, left shoe image, right shoe image, right leg image, top image, Multiple sub-images such as the right arm image and the glasses image. Using a classifier or a regressor to analyze each sub-image, different user attribute information can be obtained. For example, classifying hair images can obtain the user's hairstyle style and hair quality; analyzing facial images can obtain attribute information such as the user's gender, age, race, expression, skin condition, facial features, etc. , right arm image, left leg image, right leg image analysis can obtain other features such as user's strength and fitness, and analysis of bag image, glasses image, left shoe image, right shoe image, skirt image and top image can Obtain attribute information such as the user's preferred clothing and accessories types, user's preferred brand, price, and matching products. The analysis of the glasses image can also obtain the user's preferred glasses function information. All these user attribute information can be represented quantitatively, for example, can be represented in a hierarchical manner as described above.

Returning to FIG. 1 , further, performing the group analysis on the users includes grouping the users. An optional implementation manner is to use a classifier to classify the multiple users in the image according to the social relationship based on the clothing, the degree of association of postures and the relative position information of the multiple users in the image. For example, whether multiple users are in a relationship of a couple or family member can be determined based on whether the clothing styles and styles of the multiple users in the image are the same, and whether the users are in a relationship of friends or a relationship of lovers can also be analyzed according to the degree of intimacy of the users. For example, when it is detected that the two users have physical contact, it may be preliminarily determined that the two users are in a friend relationship or a family relationship, and then it is determined whether the two users are in a relationship based on whether the clothes of the two users are the same. Another optional implementation is to use user attribute information to cluster multiple users in the image based on the individual analysis results. The clustering method may be to quantify the attribute information of each user, calculate the distance between the user attribute information, group the user attribute information whose distance is less than a preset threshold into a group, and then group the corresponding users. Optionally, during clustering, group attribute information (eg, social relationship) of users may be preferentially used for clustering, and then individual user attribute information may be used for clustering.

Through the acquisition method of user attribute information described above, not only richer shallow features can be obtained, such as gender, age, race, health status, makeup style, accessory style, etc., but also the deep features of users, such as personality , purchasing power, etc., so that content that is more in line with user needs or more interesting to users can be recommended to users based on these characteristics, thereby improving the conversion rate of the recommended content.

In some embodiments, the manner of acquiring the environmental attribute information may include, but is not limited to, receiving current time information through the network and/or receiving spatial information corresponding to the display position of the recommended content through the network. The time information may include at least one of the following: current date and time, weather conditions, festival information, and current popular events. Spatial information may include the geographic location of the display location and/or landmarks in the vicinity, such as airports, waiting rooms, commercial centers, and the like.

In the embodiment of the present application, in addition to acquiring user attributes, environmental attributes can also be analyzed and acquired, and the recommended content can be analyzed and decided by using the environmental attributes, which can provide recommended content that is more in line with the environmental state and improve the timeliness of content recommendation. sex.

In step 102, the recommended content is synthesized based on the user attribute information and/or the environment attribute information.

In many embodiments of the present application, content is divided into multiple elements, and various combinations of these elements can be performed to generate different content. In this way, the content recommendation system does not need to store a large amount of fixed content in advance, but only needs to store various content elements to generate rich content, so this kind of changing content can also be called dynamic content.

In some embodiments, step 102 may include step 1021 : determining candidate content elements based on user attribute information and/or environmental attribute information.

In this embodiment, the candidate content elements may include candidate object elements and candidate scene elements. In some optional implementation manners, the object element may be a commodity, and the scene element may be an advertisement element. Correspondingly, the candidate object may be a candidate commodity, and the candidate scene element may be a candidate advertisement element. Object elements and scene elements can have various properties. Attributes of object elements may include, but are not limited to, at least one of the category, price, color, and brand of the item, and attributes of scene elements may include, but are not limited to, visual style, storyline, suitable items, characters, time, place, and soundtrack at least one of.

In some implementations, based on the user attribute information and/or the environment attribute information obtained in step 101, a recommendation model set may be used to determine the candidate element content. The recommendation model set may include a first recommendation model set for recommending object elements and/or scene elements according to user attribute information, a second recommendation model set for jointly recommending object elements and/or scene elements, and recommending object elements and/or scene elements according to environmental attribute information. or at least one item from the third recommendation model set of scene elements.

In a further implementation, the first recommendation model set may be a set of sub-models of interest relationships between user attribute information and attributes of object elements and/or attributes of scene elements, which may include at least one attribute of user attribute information and attributes of object elements and/or submodels of interest relationships between attributes of scene elements. The second recommendation model set may be a set of sub-models of interest relationships among attributes of object elements and/or scene elements, which may include at least one sub-model of interest relationships among attributes of object elements and/or scene elements. The third recommendation model set may be a set of sub-models of interest relationships between environment attribute information and attributes of object elements and/or attributes of scene elements, which may include at least one relationship between environment attribute information and attributes of object elements and/or scene elements A submodel of the interest relationship between attributes.

Taking the object element as a commodity and the scene element as an advertising element as an example, each sub-model in the first recommendation model set can represent a mapping relationship for recommending each commodity or each advertising element according to user attributes; the second recommendation model Each sub-model in the set can represent a mapping relationship where different commodities/advertising elements are jointly recommended, and each sub-model in the third recommendation model set can represent a mapping for recommending each commodity or each advertising element according to environmental attributes relation.

In some optional implementations of this embodiment, after the recommendation model set is determined, at least one sub-model in the recommendation model set may be used to recommend content related to user attributes and environmental attributes.

Please refer to FIG. 3, which shows an exemplary flowchart of determining candidate content elements according to one embodiment of the present application. In the embodiment corresponding to FIG. 3 , the method for determining candidate content elements by using the recommendation model set may include:

In step 301, the sub-models in the recommended model set are trained based on the statistical data of the interest degree to determine the parameters of the sub-models.

As mentioned above, each recommendation model set may include a set of sub-models, and sub-models may represent user attributes and object elements, user attributes and scene elements, between different object elements, between different scene elements, and environment attributes and objects The mapping relationship between elements, environment attributes and scene elements. In this embodiment, such a mapping relationship can be derived from the statistical data of the interest degree. Specifically, each sub-model can be trained based on the statistical data of interest degree, and the parameters of the sub-model can be obtained. The statistical data of interest degree may include: statistical data of interest degree of user attribute information on attributes of object elements and/or attributes of scene elements, statistical data of interest degree between attributes of different object elements, and statistical data of interest between attributes of different scene elements. Interest degree statistics, interest degree statistics between attributes of object elements and attributes of scene elements, and interest degree statistics of attributes of object elements and/or attributes of scene elements from environmental attribute information.

Interestingness statistics can be quantified into multiple levels or as normalized values. The acquisition method can be obtained through the data statistics of online shopping websites. For example, the corresponding relationship between the number of views and purchases of a certain commodity on the online shopping website and the age group of the users who browse or purchase the commodity can be calculated, so that the statistics are different. Statistical data on the interest of users in this product by age group. For another example, the statistical data of the interest degree of refrigerators and washing machines, and the interest degree of refrigerator brands and washing machine brands can be calculated through the statistics of the number of users who purchase multiple commodities at the same time (for example, the number of users who purchase a refrigerator of a certain brand and a washing machine of another brand at the same time). Statistical data. Another way to obtain the statistical data of interest degree is to conduct statistics through questionnaires. For example, a targeted questionnaire can be designed to count the interest of users of different ages, personalities, and purchasing power in different commodity brands. It can also be set through experience, for example, the statistical data of the interest degree of women in cosmetics can be set according to experience, and normalized to 0.8, while the statistical data of the interest degree of men in cosmetics can be set to 0.2.

Table 1 shows an example of giving the statistical data of the interest degree of age in the attribute of the object element in the user attribute information in the form of a list. where the interestingness is normalized to 0 to 1.

Table 1 Statistical data table of age’s interest in brands

Disney Gap Eland ... 0-3 0.8 0.6 0.0 ... 3-5 0.6 0.5 0.0 ... 5-10 0.8 0.6 0.3 ... 10-20 0.3 0.8 0.8 ... ... ... ... ... ...

As can be seen from Table 1, the age-to-brand interest statistics table counts the interest of users of each age in product brands. Similarly, the interest between other user attribute information and different product attributes or advertising elements can be counted. , the degree of interest between different commodities, the degree of interest of different advertising elements, and the degree of interest of environmental attribute information to different commodities or different advertising elements.

Optionally, in order to enable the trained sub-model to adapt to changes in the environment, the corresponding sub-model may be updated based on new object elements and scene elements. For example, brand-related interest statistics and sub-models can be updated based on new branded items. In addition, the statistical data of the degree of interest may be updated in a certain period of time, and the corresponding sub-model is trained by using the updated statistical data of the degree of interest to obtain the updated sub-model. For example, the sub-model can be updated quarterly based on feedback from merchants.

Step 302, establishing a global energy function based on the recommended model set.

After the sub-models in the recommended model set are obtained through training, the object elements and scene elements that meet the requirements can be searched from the object element database and the scene element database according to certain rules. The following functions can be established based on the first recommendation model set, the second recommendation model set and the third recommendation model set, and the object element and the scene element can be determined based on equation (1).

productSet ^* = argmin _productSet E(productSet|models,userSet,context) (1)

Among them, productSet represents a set of object elements or scene elements, productSet* represents a set of determined candidate object elements or candidate scene elements, models represents a set of recommended models, models={model ₁ , model ₂ , model ₃ }, where model ₁ represents The first recommendation model set, model ₂ represents the second recommendation model set, and model ₃ represents the third recommendation model set. userSet represents a collection of user attribute information, context represents environmental attribute information, and E(·) represents a global energy function.

Determining the recommended content is to select the object element and/or scene element with the smallest energy function from the database. The global energy function can be defined as equation (2):

In formula (2), productSet={product _j }, userSet={user _i }, where i, j, j ₁ , j ₂ are positive integers, product _j , product _j1 , product _j2 represent object elements or scene elements, user _i represents user attribute information. α ₁ , α ₂ , and α ₃ represent weight coefficients, which can be set according to experience or obtained through training.

As shown in formula (2), the global energy function may include a first energy function E ₁ (·), a second energy function E ₂ (·), and a third energy function E ₃ (·). The first energy function may be an energy function that recommends object elements or scene elements according to user attribute information. Specifically, the first energy function may be calculated according to formula (3):

Among them, i, j, p, q are positive integers, product _j profile _p represents the p th attribute of the j th object element/scene element, user _i profile _q represents the q th attribute of the ith user, β _{(p , q)} represents the weight coefficient. The first energy function may include: based on the first recommendation model set, a recommendation probability of the attribute of the object element and/or the attribute of the scene element corresponding to the user attribute information calculated by the classifier and/or the regressor.

The second energy function may be an energy function commonly recommended by different object elements or different scene elements. Specifically, the second energy function may be calculated according to formula (4):

Among them, j ₁ , j ₂ , p, q are positive integers, product _j1 profile _p represents the p th attribute of the j ₁ th object element/scene element, product _j2 profile _q represents the j ₂ th object element/scene element The qth attribute, β _(p,q) represents the weight coefficient. The second energy function may include: based on the second recommendation model set, the probability of jointly recommending the attributes of the object elements and/or the attributes of the scene elements calculated by the classifier and/or the regressor.

The third energy function can be an energy function that recommends object elements or scene elements according to the environmental attribute information. Specifically, the third energy function can be calculated according to formula (5):

Among them, i, j, p, q are positive integers, product _j profile _p represents the p-th attribute of the j-th object element/scene element, contextprofile _q represents the q-th attribute in the environmental attribute information, γ _{(p, q )} represents the weight coefficient. The third energy function may include: based on the third recommendation model set, using a classifier and/or a regressor to calculate the recommendation probability of the attributes of the object elements and/or the attributes of the scene elements corresponding to the user attribute information.

Continuing with FIG. 3 , in step 303 , a global optimization solution is performed on the global energy function to obtain candidate content elements that make the global energy function optimal.

In this embodiment, the recommended content may be determined based on the above energy function. Specifically, the global energy function can be globally optimized and solved according to equation (1). The methods of global optimization may include optimization algorithms based on genetic algorithms, linear programming, simulated annealing, and the like. After solving productSet* in equation (1), candidate object elements and candidate scene elements can be obtained.

In some embodiments, candidate content elements may be determined based on one of the first energy function, the second energy function, and the third energy function. Recommend colorful and gorgeous clothing; based on the second energy function, refrigerators and TV sets, lipstick and eyebrow pencils, cribs and feeding bottles can be used as jointly recommended products; based on the third energy function, down jacket products can also be recommended in snowy winter. Billboards at airports recommend tourist information. In some implementations, the recommended content may be determined in combination with two or three of the first energy function, the second energy function, and the third energy function. For example, a couple's T-shirt can be recommended to couples in summer, a couple's down jacket can be recommended to couples in winter, and a couple's watch and couple's ring can be recommended to couples at the same time.

The embodiment described above in conjunction with FIG. 3 describes a method for determining candidate content elements based on user attribute information and environmental attribute information. In practical applications, when the recommended content is an advertisement, after the global energy function is globally optimized and solved, it can be Obtain a set of product sets and advertising element sets that are jointly recommended by priority.

In the method for determining candidate content elements provided in this example, multiple recommended object elements and scene elements can be selected according to the user's interest degree and tendency and/or environmental information, which can provide richer recommended content. For example, when recommending an advertisement, a variety of advertisement elements and scene elements that meet the user's needs and preferences can be obtained, which enriches the advertisement content and improves the utilization rate and delivery effect of the billboard. Moreover, it can provide more vivid advertising elements to improve user experience.

Returning to FIG. 1 , step 102 may further include step 1022 , synthesizing recommended content according to candidate object elements and candidate scene elements.

In this embodiment, after the candidate object elements and the candidate scene elements are determined in step 1021, the candidate object elements may be fused, the candidate scene elements may be combined, and the recommended content may be generated by combining the candidate object elements and the candidate scene elements. Each candidate object element may be first fused with the corresponding candidate scene element, and then multiple candidate scene elements may be combined.

In some implementations, candidate object elements and candidate scene elements may be fused based on preset rules. Referring further to FIG. 4, it shows an exemplary flowchart of a method for synthesizing recommended content according to an embodiment of the present application.

As shown in FIG. 4, in step 401, the placement index, direction index and motion track index of candidate scene elements are acquired.

In this embodiment, the scene element generally has a transparent background or a specific placement position for placing the object element. Placement indexes can be established at these specific locations to set the types of object elements that can be placed at a particular location. For example, vehicles can be placed on the road, and watches can be placed on the wrist. Further, a direction index may also be established at a specific position to indicate the orientation of the object element. For example, the orientation of the vehicle may be determined based on the direction of the road. Determine the orientation of the watch according to the posture of the upper arm of the person. Further, when the candidate scene element is a dynamic element, such as a video, a motion track index can also be established to indicate the motion direction and route of the object element. For example, a road direction index may be included in the road scene so that the vehicle travels in the direction of the road.

Before merging the candidate object elements and the candidate scene elements, the above-mentioned index information of the candidate scene elements may be obtained first, including the placement index, the direction index, and the motion track index. The acquisition method can be to directly find relevant data from the database, or to perform image analysis and video analysis on scene elements, extract the features used to place candidate object elements, determine the placement index of the scene through the trained model, and extract the scene. The position feature, direction feature and motion track feature in the element, so as to obtain the direction index and motion track index.

In step 402, the candidate object elements and the candidate scene elements are fused according to the placement index, the direction index and the motion track index to generate candidate recommended content.

When synthesizing candidate content elements, you can first place the candidate object elements in a specific position in the candidate scene elements according to the placement index, then rotate the candidate object elements according to the direction index, and then move the candidate object elements according to the motion track index to synthesize a complete Candidate recommended content.

In step 403, the candidate recommended content is screened based on the correlation between the attributes of the candidate scene elements, the correlation between the attributes of the candidate object elements, and the correlation between the attributes of the candidate scene elements and the candidate object elements, and the screened The candidate recommended contents are fused to generate recommended contents.

In this embodiment, multiple candidate recommended contents may have certain correlations, such as temporal correlation, spatial correlation, character correlation, event correlation, and attribute correlation. Multiple candidate recommended contents with strong correlation can be screened according to these correlations, and candidate recommended contents unrelated to other candidate recommended contents can be filtered out, and the selected candidate recommended contents can be merged into smooth and coherent recommended contents.

The association between the candidate recommended contents may be based on the attributes of different candidate object elements contained in the candidate recommended contents, between the attributes of the candidate object elements and the attributes of the candidate scene elements, and between the attributes of different candidate scene elements. Relevance is established. Therefore, in this embodiment, screening can be performed according to the correlation between attributes of different candidate object elements, the correlation between the attributes of candidate object elements and the attributes of candidate scene elements, and the attributes of different candidate scene elements. Contains candidate recommended content corresponding to candidate object elements or candidate scene elements.

The correlations between the attributes of the above-mentioned different candidate object elements, the correlations between the attributes of the candidate object elements and the attributes of the candidate scene elements, and the correlations between the attributes of the different candidate scene elements can be obtained by the method of model training, or according to Experience manual setting. The representation of the association can be a quantified numerical value. Optionally, the attributes of the candidate object elements and the attributes of the candidate scene elements can be vectorized, and then the distance between the attributes is calculated, and the smaller the distance, the stronger the correlation. The correlation between the attributes of the object elements contained in all the candidate recommended contents and the properties of the scene elements is calculated, and the candidate recommended contents with the weakest or no correlation with other candidate recommended contents can be filtered out according to the correlation.

After screening, multiple candidate recommended contents with strong correlation can be obtained, and the screened candidate recommended contents can be connected in series according to the time sequence, spatial position relationship or event status, and the completed recommended contents can be generated.

It can be understood that, if only one candidate recommended content is obtained in step 1021, or there is no correlation between multiple candidate recommended contents, one candidate recommended content may be used as the recommended content.

For example, when recommending an advertisement, the correlation between the product element and the advertisement element can be calculated. For example, if two advertisement elements with similar video styles are highly correlated, the advertisement containing the two advertisement elements can be placed in the same advertisement. in the ad. Another example is that the scenes of two advertising elements are the same scene, the time attributes are morning and noon respectively, and the products included are cars and watches, respectively. The advertisements are chained together to form a continuous video advertisement.

It should be noted that, in the exemplary implementation of the method for synthesizing recommended content described above in conjunction with FIG. 4 , a screening step may be performed first, based on the correlation between the attributes of the candidate scene elements, the correlation between the attributes of the candidate object elements, and the candidate The correlation between the attributes of the scene elements and the candidate object elements filters the candidate object elements and the candidate scene elements, and then obtains the placement index, direction index and motion track index of the candidate scene elements, and then selects the candidate object elements and candidate object elements with high correlation. The scene elements are fused into candidate recommended content, and the candidate recommended content is connected in series according to the time attribute of the candidate scene elements to form a complete recommended content.

The content recommendation method provided by the above embodiments of the present application determines candidate object elements and candidate scene elements based on user attribute information and environment attribute information, and then synthesizes recommended content according to the candidate object elements and candidate scene elements. More information can be provided in the recommended content, and the utilization rate of the placement of the recommended content can be improved. Moreover, more targeted personalized content can be provided, thereby increasing the conversion rate of the recommended content.

The methods provided by the above embodiments can be used in an intelligent advertisement recommendation system. The system can obtain the image in front of the billboard through the camera set on the billboard, perform human body detection on the image, detect the target user A, and then determine that the target user A is paying attention to the content on the billboard through focus detection. The system can conduct individual analysis on the target user A. The analysis result is male, 40-50 years old, wearing high-end dark blue suits and black leather shoes, and the facial characteristics analysis of his personality is responsible, confident, emotionally stable, more introverted, and strong in purchasing power , you can recommend black business cars, dark high-end polo shirts, and high-end watches of a certain brand to the target user, and the recommended advertising elements can include classical music, high-end home scenes, business office scenes, urban road conditions, etc. Finally, based on the correlation between these products and advertising elements, the generated advertisement can be a video of the male protagonist wearing a blue high-end polo shirt, wearing a high-end watch, and driving a black business car to a business office building in the early morning. During the period, you can also intersperse the plot of the male protagonist watching the watch and attending the meeting and driving away in the sunset.

The intelligent advertisement recommendation system can also recommend advertisements for user groups including multiple users. For example, 6 target users A, B, C, D, E, and F can be detected by analyzing the image, and it can be determined through focus detection that the 6 target users are paying attention to the content on the billboard. The system can first perform individual analysis on 6 target users, and then perform group analysis based on the individual analysis results and the posture and relative position relationship between the 6 target users. The analysis result is that A, B, and C are a family of three, and they are regarded as user group 1, and D and E have a higher probability of being a couple, so they are regarded as user group 2, and F is regarded as user group 3. The system can generate three segments of advertisements, and recommend advertisements for user groups 1, 2, and 3 respectively.

In some embodiments, the above content recommendation method may further include:

In step 103, the recommended content is displayed by time-division presentation or space-division presentation.

If the number of users or user groups as service objects acquired in step 101 is multiple, after the recommended content is generated, the recommended content needs to be displayed to the multiple users or user groups as service objects in an appropriate manner. Taking the display of the recommended content on the electronic display screen as an example, in this embodiment, the recommended content may be displayed in the manner of time-division presentation or space-division presentation. The time-division presentation method is suitable for electronic screens with a small screen area, and the space-division presentation method is suitable for screens with a large screen area or curved screens.

In some implementations, displaying the recommended content by means of space-division presentation can be performed by: firstly dividing the display position of the recommended content into sub-regions equal to the number of users or user groups, The recommended content of a group of users is displayed in the sub-area where the focus of the user/group of users is located. It should be noted that pupil position detection and depth detection may be performed on the face image during the focus detection of the user's line of sight, thereby determining the position of the area that the user pays attention to on the screen. Further, the user's field of view can also be determined according to the position of the user's pupil, thereby determining the size of the displayed recommended content.

In a further implementation, the focus position of the user's line of sight can also be tracked, and the display position of the recommended content of each or each group of users can be adjusted according to the change of the focus position of the user's line of sight. If the user or user group is in motion, the change of their position can be tracked through pedestrian detection, or when the user or user group is stationary but the area of interest changes, the change in the focus position of the user's sight can be detected in real time based on the pupil position , so as to obtain the change status of the location that the user pays attention to. At this time, the display position of the recommended content can be adjusted, so that the content recommended for the user or user group can always be projected into the visual field of the user or user group.

Referring further to FIG. 5a, it shows a schematic diagram of an effect of displaying recommended content in a space-division presentation manner. The scenario of Figure 5a can utilize billboards in the lobby of a shopping mall or hotel lobby to recommend advertisements to customers. In FIG. 5a , the display screen for displaying recommended content is a cylindrical screen 510 . The system detects target service object users 501 and 502 , where the focus of sight of user 501 is located in area 511 , and the focus of sight of user 502 is located in area 512 . Through individual analysis of user 501 and user 502, it is concluded that user 501 has a greater interest and demand for watches, and user 502 has a greater demand for cars, so watch advertisements and cars can be displayed in areas 511 and 512 respectively. Advertising, which realizes the recommendation of different personalized advertisements for different users at different positions on the cylindrical screen.

Referring further to FIG. 5b, it shows a schematic diagram of another effect of displaying the recommended content by using the space division presentation. The display location shown in Figure 5b may be a flat screen display in a subway-like transfer corridor. These displays can be laid along the wall, and the displays on the wall can display personalized advertisements during the user's transfer process. As shown in FIG. 5b, the screen 520 may be divided into a plurality of sub-regions. The focus of sight of user 503, user 504 and user 505 is in area 521, area 522 and area 523, respectively. The content recommended to each user can be displayed on the corresponding area. For example, an advertisement of shirts and skirts is recommended to the user 503 , an advertisement of a watch is recommended to the user 504 , and an advertisement of a car is recommended to the user 505 . In addition, the focus position of the user's sight can be tracked in real time during the user's movement, and the position of the displayed recommended content can be adjusted according to the change of the focus position of the user's sight. For example, when the focus position of the user's sight moves to the area 522, the content displayed in the area 522 can be switched to a car advertisement. When the users overlap, the recommended content corresponding to the users whose sight is not obstructed can be displayed on the screen.

The screen used for time-division presentation may be a raster display screen, and different recommended contents are switched through the rapid movement of the raster.

In some implementations, displaying the recommended content in a time-division presentation manner may be performed by switching at least one recommended content corresponding to each or each group of users at a certain time interval on the display position of the recommended content. The time interval may be the visual persistence time of the human eye, so that the display of multiple recommended contents can be realized by utilizing the visual persistence phenomenon of the human eye.

In a further implementation, the focus position of the user's sight line can also be tracked, and the display angle of the recommended content for each or each group of users can be adjusted according to the change of the focus position of the user's sight line. While displaying the recommended content, the position of the focus and the depth information of the user can be detected in real time through focus detection, so as to determine the change of the user's field of view, and then the direction of the grating can be adjusted based on the change of the user's field of view, so that the recommended content is always displayed in the user's field of vision.

Referring further to FIG. 6 , it shows a schematic diagram of the principle of displaying recommended content in a time-division presentation manner. As shown in FIG. 6, 602 is a camera for capturing images, and a light emitting diode (LED) projection array 601 presents images or videos to the user through a grating display screen 603, wherein when the grating moves to a certain position, the left eye and the right eye A user whose eyes are located in area 611 and area 612, respectively, can see the first type of images containing the content recommended for the user; when the raster is transformed to another position, users whose left and right eyes are located in area 613 and area 614, respectively, can See a second category of images containing recommended content for that user. The camera 602 can detect changes in the focus position of the user's sight line in real time, and the grating adjusts the angle with the movement of the user's sight line to ensure that the displayed recommended content is always within the user's field of vision.

The above method of displaying recommended content through space division or time division presentation can simultaneously display multiple personalized recommended content to multiple or multiple groups of users, improve the utilization rate of the display position of the recommended content, and can automatically adjust through focus detection Placement, making content recommendation more intelligent.

Referring further to FIG. 7 , it shows a schematic structural diagram of a content recommendation apparatus according to an embodiment of the present application.

As shown in FIG. 7 , the content recommendation apparatus 700 may include an acquisition unit 701 and a synthesis unit 702 . The obtaining unit 701 may be configured to obtain user attribute information and/or environment attribute information. The synthesizing unit 702 may be configured to synthesize the recommended content based on the user attribute information and/or the environment attribute information. In some embodiments, the synthesis unit 702 may include a determination subunit 7021 and a synthesis subunit 7022 . The determining subunit 7021 is configured to determine candidate content elements based on the user attribute information and/or environment attribute information acquired by the acquiring unit 701 , where the candidate content elements may include candidate object elements and candidate scene elements. The synthesizing subunit 7022 may be configured to synthesize the recommended content according to the candidate object elements and the candidate scene elements determined by the determining subunit 7021 .

In this embodiment, the acquiring unit 701 may extract the user attribute information based on the image captured by the camera. User attribute information may include user individual attribute information and group attribute information. The individual attribute information may be information obtained by analyzing the individual characteristics of each user, and may include information such as the user's gender, age, race, clothing style, makeup style, health status, and posture. The group attribute information may be information obtained based on the relationship between multiple users, and may be the user's social relationship information, including family relationship, lover relationship, friend relationship, and the like. In some implementations, user attribute information can be retrieved through various classifiers or regressors.

In some implementations, the manner in which the obtaining unit 701 is configured for the environmental attribute information may include, but is not limited to, receiving current time information through the network and/or receiving spatial information corresponding to the display position of the recommended content through the network.

In some implementations, the determining subunit 7021 may construct a global energy function based on the user attribute information and the environmental attribute information obtained by the obtaining unit 701, and use the recommended model set to construct a global energy function, and optimize the global energy function to determine the candidate element content. The recommendation model set may include a first recommendation model set for recommending object elements and/or scene elements according to user attribute information, a second recommendation model set for jointly recommending object elements and/or scene elements, and recommending object elements and/or scene elements according to environmental attribute information. or at least one item from the third recommendation model set of scene elements.

In a further implementation, the first recommendation model set may be a set of sub-models of interest relationships between user attribute information and attributes of object elements and/or attributes of scene elements, which may include at least one attribute of user attribute information and attributes of object elements and/or submodels of interest relationships between attributes of scene elements. The second recommendation model set may be a set of sub-models of interest relationships among attributes of object elements and/or scene elements, which may include at least one sub-model of interest relationships among attributes of object elements and/or scene elements. The third recommendation model set may be a set of sub-models of interest relationships between environment attribute information and attributes of object elements and/or attributes of scene elements, which may include at least one relationship between environment attribute information and attributes of object elements and/or scene elements A submodel of the interest relationship between attributes. The above interest relationship can be represented by interest degree statistical data. The methods of obtaining the statistical data of interest degree include but are not limited to: experience setting, questionnaire statistics and online shopping website data statistics.

In some implementations, the synthesis subunit 7022 can obtain the placement index, direction index, and motion track index of the candidate scene element, fuse the candidate object element with the candidate scene element according to the index, and generate candidate recommended content, and then can The correlation between scene elements, the correlation between different candidate scene elements, and the correlation between different candidate object elements screen the candidate recommended content, and finally connect the selected candidate recommended content in chronological order to form a complete and smooth recommendation. content.

In some embodiments, the content recommendation apparatus may further include a presentation unit 703 . The presentation unit 703 may be configured to present the recommended content synthesized by the synthesis unit 702 in a time-division presentation or space-division presentation. The time-division presentation method can present the recommended content on the display screen with the movable grating installed. By changing the angle of the grating, the visual persistence of the human eye is used to switch the recommended content corresponding to multiple users or user groups within the visual persistence time. . The space-division presentation method can present the recommended content on a larger area screen or a curved screen, divide the screen into multiple sub-areas, and display the corresponding recommended content in the sub-areas concerned by each user or each group of users. In a further implementation, it is also possible to track changes in the implementation focus of users or groups of users, and adjust the display position or angle of the recommended content in real time.

The content recommendation apparatus provided by the above embodiments of the present application can provide more targeted personalized content, and provide more information in the recommended content, so as to improve the utilization rate of the recommended content display position and the conversion rate of the recommended content.

It should be understood that the units recorded in the content recommendation apparatus 700 correspond to the respective steps in the methods described with reference to FIGS. 1-6 . Therefore, the operations and features described above with respect to the method are also applicable to the content recommendation apparatus 700 and the units included therein, and details are not described herein again.

Referring further to FIG. 8 , it shows a schematic structural diagram of a content recommendation system according to an embodiment of the present application.

The content recommendation system 800 may include at least a processor 801 and a display device 802 . The processor 801 may include the content recommendation apparatus 700 described above in conjunction with FIG. 7 . The display device may be configured to display the recommended content generated by the processor. It can be understood that the processor may be an independent processing unit for executing the content recommendation method. In some implementations, the content recommendation system may also include an input device such as a keyboard, a mouse, etc.; a memory, such as a hard disk, for storing candidate object elements and candidate scene elements; a communication unit for a network interface card such as a LAN card, modem, etc. , performing communication processing via a network such as the Internet; and a removable medium such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, so that a computer program read therefrom can be installed into the memory as necessary.

As another aspect, the present application also provides a computer-readable storage medium, and the computer-readable storage medium may be the computer-readable storage medium included in the apparatus described in the foregoing embodiments; A computer-readable storage medium built into a terminal device. The computer-readable storage medium stores one or more programs, which may contain program code for executing the methods shown in the flowcharts.

The flowchart and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, apparatus, methods and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code that contains one or more logic for implementing the specified logic Executable instructions for the function. It should also be noted that, in some alternative implementations, the functions noted in the blocks may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It is also noted that each block of the block diagrams and/or flowchart illustrations, and combinations of blocks in the block diagrams and/or flowchart illustrations, can be implemented in dedicated hardware-based systems that perform the specified functions or operations , or can be implemented in a combination of dedicated hardware and computer instructions.

The above description is only a preferred embodiment of the present application and an illustration of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to the technical solution formed by the specific combination of the above-mentioned technical features, and should also cover the above-mentioned technical features without departing from the inventive concept. Other technical solutions formed by any combination of its equivalent features. For example, a technical solution is formed by replacing the above-mentioned features with the technical features disclosed in this application (but not limited to) with similar functions.

Claims (54)

1. A method for recommending content, the method comprising:

acquiring user attribute information and environment attribute information;

synthesizing recommended content based on the user attribute information and the environment attribute information;

tracking a gaze focus position of a user; and

dynamically displaying the recommended content of the user at a display position corresponding to the sight focus position of the user, wherein the dynamic display comprises the following steps: determining sub-areas corresponding to users on a display screen, wherein the sub-areas corresponding to the users are different; and displaying the recommended content of the user on the sub-area of the display screen, and changing the corresponding sub-area according to the movement of the user.

2. The method of claim 1, wherein the composite recommendation comprises:

determining candidate content elements based on the user attribute information and/or the environment attribute information, the candidate content elements including candidate object elements and/or candidate scene elements; and

and synthesizing the recommended content according to the candidate object element and/or the candidate scene element.

3. The method of claim 2, wherein the obtaining user attribute information comprises:

acquiring an image of an area where the display position of the recommended content is located;

determining a user as a service object from the image; and

performing individual analysis and population analysis on the user.

4. The method of claim 3, the determining a user from the image as a service object, comprising:

detecting a pedestrian in the image and a gaze focus position of the pedestrian;

judging whether the sight focus position of the pedestrian is located at the display position of the recommended content; and

if yes, determining that the pedestrian is the user serving as the service object.

5. The method of any of claims 3-4, wherein the performing an individual analysis of the user comprises:

dividing each user in the image into a plurality of sub-images according to the human body part; and

the sub-images are analyzed using a classifier and/or a regressor to obtain user attribute information.

6. The method of any of claims 3-4, wherein the performing a group analysis on the users comprises grouping the users, comprising:

classifying the plurality of users in the image according to social relations by adopting a classifier based on the clothing, the association degree of the postures and the relative position information of the plurality of users in the image; and/or

And clustering a plurality of users in the image by utilizing the user attribute information based on the individual analysis result.

7. The method of claim 1, wherein the obtaining user attribute information comprises:

capturing images of a plurality of users of an electronic device; dividing the plurality of users into a plurality of groups based on distances between the plurality of users; and obtaining group user attribute information corresponding to each of the plurality of groups; and the composite recommended content includes: providing group recommended content corresponding to each of the plurality of groups based on the group user attribute information and the environment attribute information.

8. The method of claim 1, wherein the environment attribute information comprises:

at least one of location information, time information, weather information, holiday information, and current trending events of the electronic device.

9. The method of claim 2, wherein the determining candidate content elements comprises: determining candidate element content using a set of recommendation models based on the user attribute information and/or the environment attribute information,

wherein the set of recommendation models includes at least one of:

the method comprises the steps of recommending a first recommendation model set of object elements and/or scene elements according to user attribute information, recommending a second recommendation model set of object elements and/or scene elements which are jointly recommended, and recommending a third recommendation model set of object elements and/or scene elements according to environment attribute information.

10. The method of claim 9,

the first recommendation model set comprises at least one sub-model of interest relationship between user attribute information and attributes of object elements and/or attributes of scene elements;

the second recommendation model set comprises at least one sub-model of interest relationship between attributes of object elements and/or attributes of scene elements;

the third recommendation model set comprises at least one sub-model of interest relationship between the environment attribute information and the attributes of the object elements and/or the attributes of the scene elements.

11. The method of claim 10, wherein determining candidate content elements using a set of recommendation models based on the user attribute information and/or the environment attribute information comprises:

training a sub-model in the recommendation model set based on the interestingness statistical data to determine parameters of the sub-model;

establishing a global energy function based on the recommendation model set;

carrying out global optimization solution on the global energy function to obtain candidate content elements enabling the global energy function to be optimal;

wherein the interestingness statistics comprise:

the interest degree statistical data of the user attribute information to the attribute of the object element and/or the attribute of the scene element;

interestingness statistics between attributes of different object elements;

interestingness statistics between attributes of different scene elements;

interestingness statistics between attributes of the object elements and attributes of the scene elements; and

and the environment attribute information is used for calculating interest degree statistical data of the attribute of the object element and/or the attribute of the scene element.

12. The method of claim 11, wherein the global energy function comprises a first energy function, a second energy function, and a third energy function;

the first energy function comprises: based on the first recommendation model set, adopting a classifier and/or a regressor to calculate recommendation probabilities of the attributes of the object elements and/or the attributes of the scene elements corresponding to the user attribute information;

the second energy function comprises: based on the second recommendation model set, calculating the probability that the attributes of the object elements and/or the attributes of the scene elements are jointly recommended by adopting a classifier and/or a regressor;

the third energy function comprises: and calculating recommendation probabilities of the attributes of the object elements and/or the attributes of the scene elements corresponding to the environment attribute information by adopting a classifier and/or a regressor based on the third recommendation model set.

13. The method of claim 11, wherein the interestingness statistics are obtained by at least one of: experience setting, questionnaire statistics, and data statistics of online shopping websites.

14. The method of claim 2, wherein synthesizing recommended content from the candidate object elements and candidate scene elements comprises:

obtaining a placement index, a direction index and a motion trail index of the candidate scene element;

fusing the candidate object elements and the candidate scene elements according to the placement index, the direction index and the motion trail index to generate candidate recommended contents; and

and screening the candidate recommended contents based on the correlation among the attributes of the candidate scene elements, the correlation among the attributes of the candidate object elements and the correlation between the attributes of the candidate scene elements and the attributes of the candidate object elements, and fusing the screened candidate recommended contents to generate the recommended contents.

15. The method of claim 7, further comprising:

and displaying the recommended content in a time-division presentation or space-division presentation mode.

16. The method of claim 15, wherein the presenting the recommended content in a spatial division presentation comprises:

dividing the display positions of the recommended contents into sub-areas with the number equal to that of users/user groups; and

and displaying the recommended content corresponding to each user/group of users into the sub-area where the user/group of users sight focus position is located.

17. The method of claim 15, wherein the presenting the recommended content in a time-division presentation manner includes:

the recommended content corresponding to each/group of users is switched at certain time intervals at the presentation position of the recommended content.

18. The method of claim 17, wherein the presenting the recommended content in a time-division presentation further comprises:

tracking a gaze focus location of the user; and

and adjusting the display angle of the recommended content corresponding to each user/group of users according to the change of the sight focus position of the users.

19. The method of claim 2,

the object element comprises a commodity and the scene element comprises an advertisement element; and/or

The attributes of the object element include at least one of: the category, price, color, and brand of the goods; and/or

The attributes of the scene elements include at least one of: visual style, story line, suitable merchandise, people, time, place, and score.

20. The method of claim 7, further comprising:

the gaze focus position of each group of users is determined by: determining a gaze focus position of the group according to an average of gaze focus positions of a plurality of users within the group;

determining a display area of the screen corresponding to the gaze focus position of the group;

and displaying the group recommendation content on a display area of the screen.

21. The method of claim 1, wherein the dynamically displaying comprises: at least one of a color, a position, a shape, and a size of the display region corresponding to the viewpoint focus position of the user is adjusted.

22. The method of claim 1, wherein the obtaining environmental attribute information comprises:

the environment attribute information is received from a server or from an electronic device within a preset range of an electronic device recommending content to a user.

23. The method of claim 1, wherein the obtaining user attribute information and environment attribute information comprises:

the method includes acquiring user attribute information based on an image acquisition device acquiring an image of each of a plurality of users of an electronic apparatus providing contents to the users, and acquiring environment attribute information based on ambient environment information acquired by one or more of the image acquisition device and a sensor.

24. The method according to claim 2, wherein the user attribute information is derived based on a plurality of sub-images derived by segmenting each user in the image by human body part; wherein the user attribute information includes information corresponding to each of a plurality of body parts of each user.

25. The method according to claim 2, wherein synthesizing recommended content according to the candidate object element and/or candidate scene element comprises:

and synthesizing a plurality of pieces of recommended content according to the candidate object elements and/or the candidate scene elements.

26. The method of claim 1, wherein the user attribute information comprises at least one of: biometric information, gender, age, race, expression, skin condition, facial features, preferences, clothing style, accessory style, preferred brand, make-up style, health status, character, and purchasing power.

27. An apparatus for recommending contents, the apparatus comprising:

an acquisition unit configured to acquire user attribute information and environment attribute information; and

a synthesizing unit configured to synthesize recommended content based on the user attribute information and the environment attribute information;

a presentation unit configured to track a gaze focus position of a user; and

dynamically displaying recommended contents of a user at a display position corresponding to a sight focus position of the user, wherein the dynamic display comprises the following steps: determining sub-areas corresponding to users on a display screen, wherein the sub-areas corresponding to the users are different; and displaying the recommended content of the user on the sub-area of the display screen, and changing the corresponding sub-area according to the movement of the user.

28. The apparatus of claim 27, wherein the synthesis unit comprises:

a determining subunit configured to determine candidate content elements based on the user attribute information and/or the environment attribute information, the candidate content elements including candidate object elements and/or candidate scene elements;

and the synthesis subunit is configured to synthesize the recommended content according to the candidate object element and/or the candidate scene element.

29. The apparatus of claim 28, wherein the means for acquiring acquires the user attribute information comprises:

acquiring an image of an area where the display position of the recommended content is located;

determining a user as a service object from the image; and

performing individual analysis and population analysis on the user.

30. The apparatus according to claim 29, wherein the manner in which the acquisition unit determines the user as the service object from the image includes:

detecting a pedestrian in the image and a gaze focus position of the pedestrian;

judging whether the sight focus position of the pedestrian is located at the display position of the recommended content; and

if yes, determining that the pedestrian is the user serving as the service object.

31. The apparatus according to any one of claims 29-30, wherein the means for obtaining unit performs individual analysis on the user comprises:

dividing each user in the image into a plurality of sub-images according to the human body part; and

the sub-images are analyzed using a classifier and/or a regressor to obtain user attribute information.

32. The apparatus according to any of claims 29-30, wherein the means for obtaining performs group analysis on the users comprises means for grouping the users into groups, and the means for grouping the users comprises:

classifying the plurality of users in the image according to social relations by adopting a classifier based on the clothing, the association degree of the postures and the relative position information of the plurality of users in the image; and/or

And clustering a plurality of users in the image by using the user attribute information based on the individual analysis result.

33. The apparatus of claim 27, wherein the means for acquiring acquires the user attribute information comprises:

capturing images of a plurality of users of an electronic device; dividing the plurality of users into a plurality of groups based on distances between the plurality of users; and acquiring group user attribute information corresponding to each of the plurality of groups; and the composite recommended content includes: providing group recommended content corresponding to each of the plurality of groups based on the group user attribute information and the environment attribute information.

34. The apparatus of claim 27, wherein the environment attribute information comprises:

at least one of location information, time information, weather information, holiday information, and current trending events of the electronic device.

35. The apparatus of claim 28, wherein the determining the manner in which the sub-unit determines the candidate content elements comprises: determining candidate element content using a set of recommendation models based on the user attribute information and/or the environment attribute information,

wherein the set of recommendation models includes at least one of:

recommending a first recommendation model set of object elements and/or scene elements according to the user attribute information, recommending a second recommendation model set of object elements and/or scene elements which are jointly recommended, and recommending a third recommendation model set of object elements and/or scene elements according to the environment attribute information.

36. The apparatus of claim 35,

the first recommendation model set comprises at least one sub-model of interest relationship between user attribute information and attributes of object elements and/or attributes of scene elements;

the second recommendation model set comprises at least one sub-model of interest relationship between attributes of object elements and/or attributes of scene elements;

the third recommendation model set comprises at least one sub-model of interest relationship between the environment attribute information and the attributes of the object elements and/or the attributes of the scene elements.

37. The apparatus of claim 36, wherein the means for determining the candidate content elements using the set of recommendation models based on the user attribute information and/or the environment attribute information comprises:

training the submodels in the recommendation model set based on the interestingness statistical data to determine parameters of the submodels;

establishing a global energy function based on the recommendation model set;

carrying out global optimization solution on the global energy function to obtain candidate content elements enabling the global energy function to be optimal;

wherein the interestingness statistics comprise:

the interest degree statistical data of the user attribute information to the attribute of the object element and/or the attribute of the scene element;

interestingness statistics between attributes of different object elements;

interestingness statistics between attributes of different scene elements;

interestingness statistics between attributes of the object elements and attributes of the scene elements; and

and the environment attribute information is interest statistical data of the attributes of the object elements and/or the attributes of the scene elements.

38. The apparatus of claim 37, wherein the global energy function comprises a first energy function, a second energy function, and a third energy function;

the first energy function comprises: based on the first recommendation model set, adopting a classifier and/or a regressor to calculate recommendation probabilities of the attributes of the object elements and/or the attributes of the scene elements corresponding to the user attribute information;

the second energy function comprises: based on the second recommendation model set, calculating the probability that the attributes of the object elements and/or the attributes of the scene elements are jointly recommended by adopting a classifier and/or a regressor;

the third energy function comprises: and calculating recommendation probabilities of the attributes of the object elements and/or the attributes of the scene elements corresponding to the environment attribute information by adopting a classifier and/or a regressor based on the third recommendation model set.

39. The apparatus of claim 37, wherein the interestingness statistic is obtained by at least one of: experience settings, questionnaire statistics, and data statistics of online shopping websites.

40. The apparatus of claim 28, wherein the means for synthesizing the recommended content according to the candidate object element and the candidate scene element comprises:

obtaining a placement index, a direction index and a motion trail index of the candidate scene element;

fusing the candidate object elements and the candidate scene elements according to the placement index, the direction index and the motion trail index to generate candidate recommended contents; and

and screening the candidate recommended contents based on the correlation among the attributes of the candidate scene elements, the correlation among the attributes of the candidate object elements and the correlation between the attributes of the candidate scene elements and the attributes of the candidate object elements, and fusing the screened candidate recommended contents to generate the recommended contents.

41. The apparatus of claim 33, wherein the presentation unit is configured to present the recommended content in a time-division presentation or a space-division presentation.

42. The apparatus of claim 41, wherein the means for presenting the recommended content in a spatial division presentation comprises:

dividing the display positions of the recommended contents into sub-areas with the number equal to that of users/user groups; and

and displaying the recommended content corresponding to each user/group of users into the sub-area where the user/group of users sight focus positions are located.

43. The apparatus of claim 41, wherein the presenting unit presents the recommended content in a time-division presentation manner, and comprises:

the recommended content corresponding to each/group of users is switched at certain time intervals at the presentation position of the recommended content.

44. The apparatus of claim 43, wherein the means for presenting presents the recommended content in a time-division manner further comprises:

tracking a gaze focus position of the user; and

and adjusting the display angle of the recommended content corresponding to each user/group of users according to the change of the sight focus position of the users.

45. The apparatus of claim 28,

the object element comprises a commodity and the scene element comprises an advertisement element; and/or

The attributes of the object element include at least one of: the category, price, color, and brand of the good; and/or

The attributes of the scene element include at least one of: visual style, story line, suitable merchandise, people, time, place, and score.

46. The apparatus of claim 33, further comprising:

the gaze focus position of each group of users is determined by: determining a gaze focus position of the group according to an average of gaze focus positions of a plurality of users within the group;

determining a display area of the screen corresponding to the gaze focus position of the group;

and displaying the group recommendation content on a display area of the screen.

47. The apparatus of claim 27, wherein the dynamic display comprises: adjusting at least one of a color, a position, a shape, and a size of a display area corresponding to a viewpoint focus position of a user.

48. The apparatus of claim 27, wherein the obtaining unit is further configured to:

the environment attribute information is received from a server or from electronic devices within a preset range of the electronic device recommending the content to the user.

49. The apparatus of claim 27, wherein the obtaining unit is further configured to:

the method includes acquiring user attribute information based on an image acquisition device acquiring an image of each of a plurality of users of an electronic apparatus providing content to the users, and acquiring environment attribute information based on ambient environment information acquired by one or more of the image acquisition device and a sensor.

50. The apparatus of claim 28, wherein the user attribute information is derived based on a plurality of sub-images obtained by segmenting each user in the image by human body part; wherein the user attribute information includes information corresponding to each of a plurality of body parts of each user.

51. The apparatus of claim 28, wherein the synthesis subunit is further configured to:

and synthesizing a plurality of pieces of recommended content according to the candidate object elements and/or the candidate scene elements.

52. The apparatus of claim 27, wherein the user attribute information comprises at least one of: biometric information, gender, age, race, expression, skin condition, facial features, preferences, clothing style, accessory style, preferred brand, make-up style, health status, character, and purchasing power.

53. A content recommendation system, characterized in that the system comprises a processor and a display device;

the display device is configured to display recommended content;

the processor comprises a content recommendation device according to any of claims 27-52.

54. An electronic device comprising a processor and a display device;

the display device is configured to display recommended content;

the processor is configured to perform the method of any of claims 1-26.

Images