TWI901377B - Information recommendation system and method - Google Patents

Abstract

An information recommendation system includes a processing unit and a storage unit electrically connected to the processing unit and storing an information database containing multiple candidate information and corresponding information feature data. The processing unit is configured to: perform a vector conversion process on a requirement description to generate requirement feature data in vector form; calculate the correlation between each piece of information feature data and the requirement feature data based on multiple information feature vectors contained in each piece of information feature data; select the most matching piece of information feature data from the information feature data; and transmit the candidate information corresponding to the most matching piece of information feature data to a user electronic device corresponding to the requirement description data.

Description

Information recommendation system and method

The present invention relates to a recommendation system, and more particularly to an information recommendation system for recommending information via the Internet. The present invention also relates to an information recommendation method for recommending information via the Internet.

In modern society, many businesses use automated recommendation systems to recommend various digital information to users, such as digital articles, advertising copy, and product introductions. Therefore, how to further improve the accuracy of recommendation systems has become a question worth exploring.

Therefore, one of the objectives of the present invention is to provide an information recommendation system that helps improve the accuracy of recommendations.

The information recommendation system of the present invention includes a processing unit and a storage unit electrically connected to the processing unit. The storage unit stores an information database containing a plurality of candidate information and a plurality of information feature data corresponding to each candidate information. Each piece of information feature data includes a plurality of information feature vectors, and each information feature vector represents a characteristic of the candidate information to which the information feature data corresponds. The processing unit is used to: after obtaining a requirement description data, perform a vector conversion process on the requirement description data to generate requirement feature data presented in vector form and representing the characteristics of the requirement description data; calculate the correlation between each information feature data and the requirement feature data based on the information feature vectors of each information feature data, and use the information feature data with the highest correlation with the requirement feature data among the information feature data as the most matching information feature data; and transmit the candidate information corresponding to the most matching information feature data to a user-end electronic device corresponding to the requirement description data.

In some embodiments of the information recommendation system of the present invention, the demand description data is an instant text message transmitted by the user electronic device, and the demand feature data is a semantic vector used to describe the semantics of the instant text message. Moreover, for each information feature data, the processing unit calculates multiple similarity scores corresponding to each information feature vector and the semantic vector of the information feature data, and then calculates the correlation between the information feature data and the semantic vector based on the average value of the similarity scores, wherein each similarity score represents the similarity between the information feature vector corresponding to the similarity score and the semantic vector.

In some implementations of the information recommendation system of the present invention, after selecting the best-matching information feature data, the processing unit also determines whether the correlation between the best-matching information feature data and the semantic vector is greater than or equal to a correlation threshold value, and only when the processing unit determines that the correlation between the best-matching information feature data and the semantic vector is greater than or equal to the correlation threshold value does it transmit the candidate information corresponding to the best-matching information feature data to the user-end electronic device.

In some embodiments of the information recommendation system of the present invention, the storage unit further stores threshold data, which includes multiple threshold parameters and multiple topic type data corresponding to the threshold parameters. After selecting the most matching information feature data, the processing unit further selects a most matching topic type data from the topic type data that has the highest semantic match with the instant text message, and uses the value of the threshold parameter corresponding to the most matching topic type data as the relevance threshold.

In some embodiments of the information recommendation system of the present invention, the demand description data is customer attribute data contained in a customer database and corresponding to the user-end electronic device, and the demand feature data includes multiple preference feature vectors for describing customer preferences. Moreover, for each information feature data, the processing unit calculates multiple similarity scores corresponding to one of the information feature vectors and one of the preference feature vectors based on the information feature vectors of the information feature data and the preference feature vectors of the demand feature data, and then calculates the correlation between the information feature data and the demand feature data based on the average value of the similarity scores, wherein each similarity score represents the similarity between the information feature vector corresponding to the similarity score and the demand feature data corresponding to the similarity score.

Another object of the present invention is to provide an information recommendation method that helps improve recommendation accuracy.

The information recommendation method of the present invention is implemented by an information recommendation system. The information recommendation system stores an information database, which includes multiple candidate information and multiple information feature data corresponding to the candidate information. Each information feature data includes multiple information feature vectors, and each information feature vector represents the characteristics of the candidate information corresponding to the information feature data. The information recommendation method includes: (A) after obtaining a requirement description data, performing a vector conversion process on the requirement description data to generate requirement feature data presented in vector form and representing the characteristics of the requirement description data; (B) based on the information feature vectors of each information feature data, calculating the correlation between each information feature data and the requirement feature data, and determining the information feature data with the highest correlation with the requirement feature data among the information feature data as the best matching information feature data; and (C) transmitting the candidate information corresponding to the best matching information feature data to a user-end electronic device corresponding to the requirement description data.

In some embodiments of the information recommendation method of the present invention, in step (A), the demand description data is an instant text message sent by the user electronic device. In step (B), the demand feature data is a semantic vector used to describe the semantics of the instant text message. For each piece of information feature data, the information recommendation system calculates multiple similarity scores corresponding to each information feature vector and the semantic vector of the information feature data, and then calculates the correlation between the information feature data and the semantic vector based on the average of the similarity scores, wherein each similarity score represents the similarity between the information feature vector corresponding to the similarity score and the semantic vector.

In some embodiments of the information recommendation method of the present invention, the method further includes, between (B) and (C): (D) determining whether the correlation between the best-matching information feature data and the semantic vector is greater than or equal to a correlation threshold. In step (C), the information recommendation system transmits the candidate information corresponding to the best-matching information feature data to the user electronic device only if the correlation between the best-matching information feature data and the semantic vector is determined to be greater than or equal to the correlation threshold.

In some embodiments of the information recommendation method of the present invention, the information recommendation system further stores threshold data, which includes multiple threshold parameters and multiple topic type data corresponding to the threshold parameters. The information recommendation method further includes, between (B) and (D): (E) selecting, from the topic type data, a best-matching topic type data that has the highest semantic match with the instant text message, and using the value of the threshold parameter corresponding to the best-matching topic type data as the relevance threshold.

In some implementations of the information recommendation method of the present invention, in step (A), the demand description data is customer attribute data contained in a customer database and corresponding to the user-end electronic device. In step (B), the demand feature data includes multiple preference feature vectors for describing customer preferences, and, for each information feature data, the information recommendation system calculates multiple similarity scores corresponding to one of the information feature vectors and one of the preference feature vectors based on the information feature vectors of the information feature data and the preference feature vectors of the demand feature data, and then calculates the correlation between the information feature data and the demand feature data based on the average value of the similarity scores, wherein each similarity score represents the similarity between the information feature vector corresponding to the similarity score and the demand feature data corresponding to the similarity score.

The utility of the present invention lies in that, after obtaining the demand description data, the information recommendation system can first convert it into the demand feature data in vector form, then calculate the correlation between the demand feature data and the information feature data corresponding to the candidate information to select the most relevant and best-matching information feature data, and then transmit the candidate information corresponding to the best-matching information feature data to the user electronic device to achieve information recommendation. In this way, even if the requirement description data only contains words that are identical or similar to the content of the candidate information, but no words that completely match the content of the candidate information, the information recommendation system can use the language processing model to select the candidate information with the highest semantic relevance to the requirement description data for recommendation. Therefore, compared to the approach of establishing a corresponding bag-of-words model for each candidate information, the present invention will not miss out on candidate information that is actually suitable for recommendation due to the lack of completely matching keywords in the requirement description data, thereby helping to improve the accuracy of information recommendations.

Before the present invention is described in detail, it should be noted that, unless otherwise specified, the term "electrically connected" as used in this patent specification is used to describe the "coupled" relationship between computer hardware (e.g., electronic systems, devices, apparatuses, units, components), and generally refers to "wired electrical connections" achieved by physically connecting multiple computer hardware components to each other through conductive/semiconductor materials, as well as "radio connections" achieved by wireless data transmission using wireless communication technologies (such as, but not limited to, wireless networks, Bluetooth, and electromagnetic induction). On the other hand, unless otherwise specified, the "electrical connection" described in this patent specification also generally refers to a "direct electrical connection" achieved by directly coupling multiple computer hardware components to each other, and an "indirect electrical connection" achieved by indirectly coupling multiple computer hardware components through other computer hardware components.

The term "unit" as used in this patent specification refers to computer hardware, not software. For example, a "processing unit" is used to refer to computer hardware that performs data processing. On the other hand, a "unit" as used in this patent specification can refer to a single piece of computer hardware that performs a specific function, or a group of computer hardware that performs similar functions. For example, a "processing unit" can refer to a single processor that performs data processing, but it can also refer to a group of processors.

Referring to FIG. 1 , an embodiment of the information recommendation system 1 of the present invention belongs to a service organization. Moreover, the information recommendation system 1 of this embodiment is suitable for use in conjunction with a user-end electronic device 5 shown in FIG. 1 .

In the application environment of this embodiment, the service institution is, for example, a commercial bank that provides financial services. However, the service institution may also be other types of business institutions that provide different services. Therefore, the application of this embodiment is not limited to the aforementioned example.

The user-end electronic device 5 is, for example, a mobile phone, tablet computer, laptop computer, or desktop computer, and belongs to a customer of the service organization. It should be noted that the information recommendation system 1 can actually be used with multiple user-end electronic devices. However, for ease of description, this embodiment will only be described using the user-end electronic device 5 shown in FIG. 1 .

In this embodiment, the information recommendation system 1 is a server device, and the information recommendation system 1 includes a processing unit 11 and a storage unit 12 electrically connected to the processing unit 11.

In this embodiment, the processing unit 11 is a processor implemented as an integrated circuit and having data processing and instruction transmission and reception functions. However, in different embodiments, the processing unit 11 may also be a circuit assembly including a processor and a circuit board, or a collection of multiple processors. The storage unit 12 is a data storage device (such as a hard drive) for storing digital data. However, in different embodiments, the storage unit 12 may also be a collection of multiple storage devices of the same or different types.

Furthermore, in other embodiments, the information recommendation system 1 may also be implemented as multiple electrically connected server devices. In this case, the processing unit 11 may be implemented as a collection of multiple processors/circuit components included in each of the server devices, and the storage unit 12 may be implemented as a collection of multiple storage devices included in each of the server devices. Based on the above, the actual implementation of the information recommendation system 1 in terms of computer hardware is not limited to this embodiment.

The storage unit 12 stores a language processing model M, an information database D1, a customer database D2, and threshold data D3.

The language processing model M is a large language model (LLM) implemented using deep learning technology and capable of natural language understanding. However, in various embodiments, the language processing model M can be any model that can convert text messages into contextual vectors using sentence embedding, word embedding, or chapter embedding, and is not limited to a large language model.

The information database D1 includes a plurality of candidate information D11 and a plurality of information feature data D12 corresponding to the candidate information D11.

Each candidate information D11 may be in the form of a text file, a still image file, a moving image file, an audio file, or a multimedia file containing multiple of these. Furthermore, the content of each candidate information D11 may include, but is not limited to, one or more of a digital article, a digital advertisement, service information, or a product introduction.

For example, two pieces of candidate information D11 from the candidate information D11 are designated as first candidate information D11A and second candidate information D11B in this embodiment (shown in FIG1 ). The first candidate information D11A may be a digital article titled "The US dollar enters a slow appreciation pattern, and commodity currencies await the development of policy and other fundamental factors." The second candidate information D11B may be a digital article titled "Export controls have limited short-term impact on the mobile phone and radio frequency industries."

Each piece of information feature data D12 includes multiple information feature vectors V12 (illustrated in FIG3 ), each of which is used to represent the characteristics of the candidate information D11 to which the information feature data D12 corresponds. More specifically, each information feature vector V12 is a 760-dimensional (but not limited to) context vector that can be processed by the language processing model M and is used to describe the content of the corresponding candidate information D11. Continuing with the previous example, the information feature data D12 corresponding to the first candidate information D11A is used as a first information feature data D12A. The first information feature data D12A, for example, includes four information feature vectors V12, and the four information feature vectors V12 are used to describe the four keywords corresponding to the first candidate information D11A, such as "Australian dollar", "Canadian dollar", "US dollar" and "monetary policy". The information feature data D12 corresponding to the second candidate information D11B is used as a second information feature data D12B. The second information feature data D12B, for example, includes three information feature vectors V12, and the three information feature vectors V12 are used to describe the three keywords corresponding to the second candidate information D11B, such as "mobile phone", "semiconductor" and "trade war".

It should be noted that the keywords corresponding to each piece of candidate information D11 can be determined, for example, by a topic model using Latent Dirichlet allocation (LDA), or by professionals in the relevant field to which each piece of candidate information D11 belongs.

The customer database D2 contains multiple customer attribute data D21 corresponding to multiple customers.

In this embodiment, each customer attribute data D21 includes, for example, a focus feature portion D211 and an asset feature portion D212.

In this embodiment, the attention feature portion D211 indicates the types of financial products that the corresponding customer has recently browsed (e.g., "high-tech," "stock index funds," "bonds," "solar energy"), the topics of online articles that the customer has browsed (e.g., "interest rate hikes," "real estate," "trade war," "U.S. debt," "inflation"), and the topics of online advertisements that the customer has clicked (e.g., "discounts," "credit," "handling fees"), but is not limited to these.

In this embodiment, the asset characteristics portion D212 indicates the type of assets (including foreign currencies and various investment-type financial products) currently held by the corresponding customer, such as "US dollars", "Japanese yen", "semiconductors", and "health care", but is not limited thereto.

The threshold data D3 includes a plurality of threshold parameters D31 and a plurality of subject type data D32 corresponding to the threshold parameters D31.

In this embodiment, the value of each threshold parameter D31 is between 0 and 1, and may be, for example, 0.5, 0.6, 0.7, etc., but is not limited thereto.

The topic category data D32, for example, respectively indicates a plurality of topic categories related to the candidate information D11. For example, the topic categories include a finance category, an e-commerce category, a digital media category, a government and public services category, and a health and medical category. More specifically, in this embodiment, each topic category data D32 is a context vector that describes the topic category corresponding to the topic category data D32 and can be processed by the language processing model M.

Specifically, the values of the threshold parameters D31 are set based on, for example, the number of candidate information D11 belonging to each topic type within the candidate information D11. For example, if a relatively large number of candidate information D11 belonging to the financial category in the information database D1 is financial, this indicates that a relatively high standard of accuracy can be applied to the recommendation of financial information. In this case, the threshold parameter D31 corresponding to the topic type data D32 indicating the financial category can be set to a relatively high value, such as 0.75. If only a small number of candidate information D11 in the information database D1 belongs to the health and medical category, this means that when recommending health and medical information, only a relatively low standard of accuracy can be required. In this case, the threshold parameter D31 corresponding to the topic type data D32 indicating the health and medical category can be set to a relatively low value, such as 0.45.

The following describes how the information recommendation system 1 of this embodiment implements an information recommendation method.

In this embodiment, the information recommendation method includes a first recommendation process and a second recommendation process. Referring to Figures 1 and 2, the following first describes how the information recommendation system 1 performs the first recommendation process.

First, in step S11, after obtaining the requirement description data corresponding to the user-end electronic device 5, the processing unit 11 uses the language processing model M to perform a vector conversion process on the requirement description data to generate requirement feature data presented in vector form and representing the characteristics of the requirement description data.

For the convenience of description, the requirement description data is referred to as a first requirement description data, and the requirement feature data is also referred to as a first requirement feature data D10 shown in FIG. 3 .

In this embodiment, the first demand description data is an instant text message sent by the user electronic device 5 according to the user's operation, and the processing unit 11 receives the first demand description data from the user electronic device 5 through the network.

For example, the first requirement description data may be, for example, "How to invest in US dollars?" or "Can I buy iPhone-related stocks?", but is not limited thereto.

Furthermore, in this embodiment, the processing unit 11 utilizes the language processing model M to perform the vector conversion process using sentence embedding to convert the first demand description data (i.e., the instant text message input by the user) into the first demand feature data D10 in vector form. More specifically, as shown in FIG3 , the first demand feature data D10 in this embodiment is a 760-dimensional semantic vector V10 (also referred to as a "semantic vector") used to describe the semantics of the instant text message, but the present invention is not limited to this.

It should be noted that, in different embodiments, the processing unit 11 may generate the semantic vector V10 by performing the vector conversion process on the first requirement description data using word embedding technology or article embedding, and is not limited to this embodiment.

After the processing unit 11 generates the semantic vector V10, the process proceeds to step S12.

In step S12, the processing unit 11 calculates the correlation between each information feature data D12 and the semantic vector V10 based on the information feature vectors V12 of each information feature data D12.

In this embodiment, for each piece of information feature data D12, the processing unit 11 first calculates a plurality of similarity scores corresponding to each information feature vector V12 and the semantic vector V10 of the information feature data D12. Next, the processing unit 11 calculates the correlation between the information feature data D12 and the semantic vector V10 based on the average of these similarity scores.

Each similarity score represents the similarity between the information feature vector V12 and the semantic vector V10 to which it corresponds. Furthermore, in this embodiment, the value of each similarity score is, for example, between 0 and 1. In this embodiment, the processing unit 11 calculates the similarity scores using a cosine similarity algorithm. However, in different embodiments, the processing unit 11 may also calculate the similarity scores using, for example, Euclidean distance, Manhattan distance, or other mathematical methods that can be used to evaluate the similarity between two vectors, and is not limited to this embodiment.

3 , taking the first information feature data D12A as an example, the processing unit 11 first calculates the similarity scores of the four information feature vectors V12 corresponding to the first information feature data D12A. The four similarity scores are used to represent the similarity between the information feature vector V12 and the semantic vector V10 describing the keyword "Australian dollar" (for example, 0.43), the similarity between the information feature vector V12 and the semantic vector V10 describing the keyword "Canadian dollar" (for example, 0.49), the similarity between the information feature vector V12 and the semantic vector V10 describing the keyword "US dollar" (for example, 0.83), and the similarity between the information feature vector V12 and the semantic vector V10 describing the keyword "monetary policy" (for example, 0.72). Then, the processing unit 11 takes the average of the four similarity scores to the third decimal place as the correlation degree between the first information feature data D12A and the semantic vector V10 (for example, 0.617).

After the processing unit 11 calculates the correlation between each information feature data D12 and the semantic vector V10, the process proceeds to step S13.

In step S13, the processing unit 11 selects the information feature data D12 having the highest correlation with the semantic vector V10 from the information feature data D12, and uses the selected information feature data D12 as the best matching information feature data.

Continuing with the previous example, assuming that the first information feature data D12A is the one with the highest correlation with the semantic vector V10 among all the information feature data D12, the processing unit 11 will select the first information feature data D12A as the most matching information feature data.

After the processing unit 11 selects the most matching information feature data, the process proceeds to step S14.

In step S14, the processing unit 11 selects a most matching topic type data item from the topic type data item D32 that has the highest semantic match with the instant text message, and uses the value of the threshold parameter D31 corresponding to the most matching topic type data item as a relevance threshold in this embodiment. In other words, in this embodiment, the processing unit 11 determines the relevance threshold based on the topic type item that is most similar to the instant text message among the topic types indicated by the topic type data item D32.

More specifically, in this embodiment, the processing unit 11, for example, utilizes the language processing model M to calculate the similarity between each topic type data D32 and the semantic vector V10 using the cosine similarity algorithm, and uses the topic type data D32 with the highest similarity to the semantic vector V10 as the most matching topic type data, but is not limited to this.

After the processing unit 11 determines the correlation threshold, the process proceeds to step S15.

In step S15, the processing unit 11 determines whether the correlation between the best-matching information feature data and the semantic vector V10 is greater than or equal to the correlation threshold. If the processing unit 11 determines yes, the process proceeds to step S16. If the processing unit 11 determines no, the process proceeds to step S17.

In step S16, following step S15, if the processing unit 11 determines that the correlation between the best-matching information feature data and the semantic vector V10 is greater than or equal to the correlation threshold, it indicates that the candidate information D11 corresponding to the best-matching information feature data has a relatively high probability of meeting the user's needs expressed in the instant text message. In this case, the processing unit 11 transmits the candidate information D11 corresponding to the best-matching information feature data to the user electronic device 5. The first recommendation process then ends.

In step S17, following step S15, if the processing unit 11 determines that the correlation between the best-matching information feature data and the semantic vector V10 is not greater than or equal to the correlation threshold, it indicates that the probability that the candidate information D11 corresponding to the best-matching information feature data meets the user's needs is insufficient. In this case, the processing unit 11 generates a no-matching information notification and transmits it to the user electronic device 5. The first recommendation process then ends.

It is worth mentioning that, in this embodiment, the values of the threshold parameters D31 are set based on the number of candidate information D11 belonging to each topic type in the candidate information D11, and the processing unit 11 determines the relevance threshold in real time based on the topic type closest to the instant text message. Therefore, the information recommendation system 1 of this embodiment can automatically adjust the judgment criteria for whether to provide candidate information D11 to the user based on the semantics expressed in the instant text message. However, in some embodiments, the relevance threshold can also be a single fixed value that is preset, and does not need to be determined by the processing unit 11 based on the semantics of the instant text message. Therefore, the state of the relevance threshold is not limited to this embodiment.

The above is an explanation of how the information recommendation system 1 executes the first recommendation process.

Referring to FIG. 1 and FIG. 4 , the following describes how the information recommendation system 1 performs the second recommendation procedure in the information recommendation method.

First, in step S21, after obtaining another requirement description data corresponding to the user-end electronic device 5, the processing unit 11 uses the language processing model M to perform the vector conversion processing on the requirement description data to generate another requirement feature data presented in vector form and representing the characteristics of the requirement description data.

For the convenience of description, the other requirement description data is referred to as a second requirement description data, and the other requirement feature data is also referred to as a second requirement feature data D20 shown in FIG. 5 .

In this embodiment, unlike the first demand description data, the second demand description data is a target customer attribute data entry corresponding to the user electronic device 5 in the customer attribute data D21 of the customer database D2. Specifically, the customer corresponding to the target customer attribute data is a target recommendation object, and the user electronic device 5 is, for example, owned by the target recommendation object. Furthermore, in this embodiment, the processing unit 11 automatically reads the storage unit 12 to obtain the second demand description data from the customer database D2, for example, when the current time reaches a pre-set automatic recommendation time, but the present invention is not limited to this.

Furthermore, referring to FIG. 5 , unlike the first demand profile data D10, the second demand profile data D20 includes multiple preference feature vectors V20 used to describe customer preferences. Each preference feature vector V20 has a dimension of 760, but is not limited to this. Specifically, in this embodiment, these preference feature vectors V20 are used to describe the characteristics of the second demand profile data, namely, the types of financial products recently viewed by the target recommendation recipient, the topics of online articles viewed, the topics of online ads clicked, and/or the types of assets currently held.

For example, as shown in FIG5 , the preference eigenvectors V20 are respectively a first preference eigenvector V20A, a second preference eigenvector V20B, a third preference eigenvector V20C, a fourth preference eigenvector V20D, and a fifth preference eigenvector V20E. Among them, the first preference feature vector V20A is used to describe that the target recommendation object recently browsed the financial product "stock index fund", the second preference feature vector V20B is used to describe that the target recommendation object recently browsed an online article with the theme of "interest rate hike", the third preference feature vector V20C is used to describe that the target recommendation object recently clicked on the online advertisement of "credit", the fourth preference feature vector V20D is used to describe that the target recommendation object currently holds "US dollar" foreign currency assets, and the fifth preference feature vector V20E is used to describe that the target recommendation object currently holds "semiconductor" type investment financial products, but it is not limited to this.

After the processing unit 11 generates the second demand characteristic data D20, the process proceeds to step S22.

In step S22, the processing unit 11 calculates the correlation between each information feature data D12 and the second demand feature data D20 according to the information feature vectors V12 of each information feature data D12.

In this embodiment, for each piece of information feature data D12, the processing unit 11 first calculates a plurality of similarity scores corresponding to one of the information feature vectors V12 and one of the preference feature vectors V20 based on the information feature data D12 and the preference feature vectors V20 of the second demand feature data D20 using a cosine similarity algorithm (but not limited to this algorithm). The processing unit 11 then calculates the correlation between the information feature data D12 and the second demand feature data D20 based on the average of these similarity scores.

Each similarity score represents the similarity between the information feature vector V12 corresponding to the similarity score and the second demand feature data D20 corresponding to the similarity score, and the value of each similarity score is, for example, between 0 and 1, but is not limited thereto.

Referring to FIG. 5 , taking the three information feature vectors V12 used to describe "mobile phone," "semiconductor," and "trade war" in the second information feature data D12B as an example, the processing unit 11 calculates five similarity scores between the information feature vector V12 describing "mobile phone" and each of the first through fifth preference feature vectors V20A through V20E. Therefore, these five similarity scores not only correspond to the information feature vector V12 describing "mobile phone," but also to the first through fifth preference feature vectors V20A through V20E in the second demand feature data D20.

Similarly, for the information feature vector V12 used to describe "semiconductor," the processing unit 11 calculates five additional similarity scores between it and each of the five preference feature vectors V20 (i.e., the first through fifth preference feature vectors V20A through V20E). Therefore, in addition to corresponding to the information feature vector V12 used to describe "semiconductor," the five additional similarity scores also correspond to the first through fifth preference feature vectors V20A through V20E.

Similarly, for the information feature vector V12 used to describe the "trade war", the processing unit 11 will calculate five more similarity scores between it and the five preference feature vectors V20. The five more similarity scores all correspond to the information feature vector V12 used to describe the "trade war" and also correspond to the first preference feature vector V20A to the fifth preference feature vector V20E, respectively.

Therefore, in this example, for the second information feature data D12B, the processing unit 11 calculates a total of fifteen similarity scores based on the three information feature vectors V12 of the second information feature data D12B and the first through fifth preference feature vectors V20A to V20E of the second demand feature data D20. Then, similar to the first recommendation process, the processing unit 11 averages these fifteen similarity scores to the third decimal place, which serves as the correlation between the second information feature data D12B and the second demand feature data D20.

After the processing unit 11 calculates the correlation between each information feature data D12 and the second demand feature data D20, the process proceeds to step S23.

In step S23, the processing unit 11 selects the information feature data D12 having the highest correlation with the second demand feature data D20 from the information feature data D12, and uses the selected information feature data D12 as another best matching information feature data.

Continuing with the previous example, assuming that the second information feature data D12B is the one with the highest correlation with the second demand feature data D20 among all the information feature data D12, the processing unit 11 will select the second information feature data D12B as the other best matching information feature data.

After the processing unit 11 selects the other best matching information feature data, the process proceeds to step S24.

In step S24, the processing unit 11 determines whether the correlation between the other best-matching information feature data and the second information feature data D12B is greater than or equal to another correlation threshold. The other correlation threshold may be, for example, the value of the threshold parameter D31 corresponding to the subject category data D32 of the digital financial service type, but is not limited thereto. If the processing unit 11 determines yes, the process proceeds to step S25. If the processing unit 11 determines no, the process proceeds to step S26.

In step S25, following step S24, if the processing unit 11 determines that the correlation between the other best-matching information feature data and the second information feature data D12B is greater than or equal to the other correlation threshold, it indicates that the candidate information D11 corresponding to the other best-matching information feature data has a relatively high probability of meeting the target recommendation subject's preferences. In this case, the processing unit 11 transmits the candidate information D11 corresponding to the other best-matching information feature data to the user electronic device 5 for viewing by the target recommendation subject. The second recommendation process then ends.

In step S26, following step S24, if the processing unit 11 determines that the correlation between the other best-matching information feature data and the second information feature data D12B is not greater than or equal to the other correlation threshold, it indicates that the probability that the candidate information D11 corresponding to the other best-matching information feature data meets the preferences of the target recommendation object is not high enough. In this case, the processing unit 11 does not transmit any candidate information D11 to the user electronic device 5, and the second recommendation process ends.

It should be noted that, in an actual implementation, the processing unit 11 may actually obtain multiple target customer attribute data from the customer database D2 at one time in step S21 as multiple second demand feature data D20, and execute steps S22 to S26 based on each second demand feature data D20.

In addition, in a similar embodiment, the information recommendation system 1 can also start executing the second recommendation process based on the local or remote operation of the user (such as the staff of the service agency).

The above is an explanation of how the information recommendation system 1 executes the second recommendation process.

It should be noted that steps S11 to S17, steps S21 to S26 and the flowcharts of FIG. 2 and FIG. 4 of this embodiment are only used to illustrate one possible implementation of the present invention. It should be understood that even if steps S11 to S17 and steps S21 to S26 are merged, split, or reordered, if the process after the merger, split, or reordering performs similar functions in a similar manner and obtains similar results compared to the present embodiment, it still falls within the applicable scope of the present invention. Therefore, steps S11 to S17 and steps S21 to S26 of the present embodiment and the flowcharts of Figures 2 and 4 are not intended to limit the applicable scope of the present invention.

In summary, by executing the information recommendation method, the information recommendation system 1 can, after obtaining the demand description data, first convert it into the demand feature data in vector form, and then calculate the correlation between the demand feature data and the information feature data D12 corresponding to the candidate information D11 to select the best matching information feature data with the highest correlation, and transmit the candidate information D11 corresponding to the best matching information feature data to the user-end electronic device 5 to achieve information recommendation. In this way, even if the demand description data only contains words that are identical or similar to the content of the candidate information D11, but no words that completely match the content of the candidate information D11, the information recommendation system 1 can use the language processing model M to select the candidate information D11 with the highest semantic relevance to the demand description data for recommendation. Therefore, compared with the method of establishing a corresponding bag-of-words model for each candidate information D11, this embodiment will not miss the candidate information D11 that is actually suitable for recommendation due to the lack of completely matching keywords in the demand description data, which helps to improve the accuracy of information recommendation and can indeed achieve the purpose of this invention.

However, the above description is merely an example of the present invention and should not be used to limit the scope of the present invention. All simple equivalent changes and modifications made within the scope of the patent application and the contents of the patent specification of the present invention are still within the scope of the present patent.

1: Information Recommendation System 11: Processing Unit 12: Storage Unit M: Language Processing Model D1: Information Database D11: Candidate Information D11A: First Candidate Information D11B: Second Candidate Information D12: Information Feature Data D12A: First Information Feature Data D12B: Second Information Feature Data V12: Information Feature Vector D2: Customer Database D21: Customer Attribute Data D211: Attention Feature Data D212: Asset Feature Data D3: Threshold Data D31: Threshold Parameter D32: Topic Type Data 5: User Electronic Device D10: First Requirement Feature Data D20: Second Requirement Feature Data V10: Semantic Vector V20: Preference eigenvector V20A: First preference eigenvector V20B: Second preference eigenvector V20C: Third preference eigenvector V20D: Fourth preference eigenvector V20E: Fifth preference eigenvector S11-S17: Steps S21-S26: Steps

Other features and functions of the present invention will be clearly presented in the embodiments with reference to the drawings.

FIG1 is a block diagram illustrating an exemplary embodiment of the information recommendation system of the present invention.

FIG2 is a flow chart for exemplarily illustrating how the embodiment implements a first recommendation process in an information recommendation method.

FIG3 is a schematic diagram exemplarily showing a piece of information feature data and a piece of first demand feature data used in the execution of the first recommendation process of the embodiment.

FIG4 is a flow chart for exemplarily illustrating how this embodiment implements a second recommendation process in an information recommendation method.

FIG5 is a schematic diagram exemplarily showing a piece of information feature data and a piece of second demand feature data used in the execution of the second recommendation process of the embodiment.

1: Information recommendation system

11: Processing unit

12: Storage unit

M: Language Processing Model

D1: Information Database

D11: Candidate Information

D11A: First Candidate Information

D11B: Second candidate information

D12: Information Feature Data

D12A: First Information Feature Data

D12B: Second information feature data

D2: Customer Database

D21: Customer attribute data

D211: Focus on the characteristic part

D212: Asset Characteristics

D3: Threshold Data

D31: Threshold value parameter

D32: Subject type data

5: Using end electronic devices

Claims (2)

An information recommendation system includes: a processing unit; and a storage unit electrically connected to the processing unit and storing an information database and a threshold value data, wherein the information database includes a plurality of candidate information and a plurality of information feature data corresponding to the candidate information, each information feature data includes a plurality of information feature vectors, and each information feature vector represents a feature of the candidate information corresponding to the information feature data, the threshold value data includes a plurality of threshold value parameters and a plurality of topic type data corresponding to the threshold value parameters, and The value of each threshold parameter is positively correlated with the number of topic types indicated by the topic type data corresponding to the threshold parameter in the waiting information; wherein the processing unit is used to: after obtaining a demand description data, perform a vector conversion process on the demand description data to generate a demand feature data presented in a vector form and representing the characteristics of the demand description data, wherein the demand description data is an instant text message sent by a user end electronic device, and the demand feature data is a semantic description of the instant text message. semantic vector; based on each information feature vector of each information feature data and the semantic vector, calculating a plurality of similarity scores respectively corresponding to the information feature vectors, and then calculating the correlation between the information feature data and the semantic vector based on the average value of the similarity scores, and taking the information feature data with the highest correlation with the semantic vector among the information feature data as the best matching information feature data, wherein each similarity score represents the similarity between the information feature vector corresponding to the similarity score and the semantic vector; from the Selecting a best-matching topic type data item from the topic type data item that has the highest semantic match with the instant text message, and using the value of the threshold parameter corresponding to the best-matching topic type data item as a relevance threshold value; determining whether the relevance between the best-matching information feature data item and the semantic vector item is greater than or equal to the relevance threshold value; and if the relevance between the best-matching information feature data item and the semantic vector item is determined to be greater than or equal to the relevance threshold value, transmitting the candidate information item corresponding to the best-matching information feature data item to the user electronic device. An information recommendation method is implemented by an information recommendation system. The information recommendation system stores an information database and threshold data. The information database includes multiple candidate information and multiple information feature data corresponding to the candidate information. Each information feature data includes multiple information feature vectors, and each information feature vector represents the characteristics of the candidate information corresponding to the information feature data. The threshold data includes multiple threshold parameters and multiple topic type data corresponding to the threshold parameters. The value of each threshold parameter is is positively correlated with the number of topic types indicated by the topic type data corresponding to the threshold value parameter in the waiting information; the information recommendation method comprises: (A) after obtaining a demand description data, performing a vector conversion process on the demand description data to generate a demand feature data presented in a vector form and representing the characteristics of the demand description data, wherein the demand description data is an instant text message sent by a user electronic device, and the demand feature data is a semantic vector for describing the semantics of the instant text message; (B) based on Based on each information feature vector and the semantic vector of each information feature data, a plurality of similarity scores corresponding to the information feature vectors are calculated, and then the correlation between the information feature data and the semantic vector is calculated based on the average value of the similarity scores, and the information feature data with the highest correlation with the semantic vector among the information feature data is used as the best matching information feature data, wherein each similarity score represents the similarity between the information feature vector corresponding to the similarity score and the semantic vector; (C) from the subject type data (D) determining whether the degree of correlation between the best-matching information feature data and the semantic vector is greater than or equal to the correlation threshold; and (E) if it is determined that the degree of correlation between the best-matching information feature data and the semantic vector is greater than or equal to the correlation threshold, transmitting the candidate information corresponding to the best-matching information feature data to the user-end electronic device.