CN111428127A - Personalized event recommendation method and system integrating topic matching and two-way preference - Google Patents

Abstract

The invention discloses a personalized event recommendation method and system integrating topic matching and bidirectional preference. First, use the document topic generation model LDA to extract the topic information of events and historical events that users participated in, and calculate the topic matching degree between users and events; The preference model of events is used to obtain user preference scores and event preference scores, respectively, to mine preference relationships more completely from the perspectives of users and events. ‑Event pair comprehensive score, and the sorted TOP‑K user‑event pairs are used as recommendation results. The performance of the recommendation algorithm of this scheme is better than that of the traditional recommendation scheme, and it can well predict the user's personalized preference, so as to achieve the purpose of personalized recommendation.

Description

Personalized event recommendation method and system integrating topic matching and bidirectional preference

technical field

The invention relates to the technical field of information recommendation, in particular to a personalized event recommendation method and system integrating topic matching and bidirectional preference.

Background technique

With the rapid development of the Internet and computer technology, traditional social networks have also developed in different directions in recent years, resulting in the formation of some special types of new social networks, such as location-based social networks (Location-Based Social Network, LBSN), a social network that forms social relationships mainly based on users’ geographic check-in information, and another complex heterogeneous social network that combines online and offline—Events-Based Social Network (EBSN), Different from the friendship established between acquaintances in traditional social networks, in event-based social networks, users establish interpersonal relationships through social activities, and users join online interest groups and offline collective social activities according to their own interests or commonalities .

Event-based social networks are in the process of rapid development. More and more users choose to participate in social activities in event-based social networks. On event-based social network platforms, users can join various online groups. The organizer or users in the group can initiate and participate in any offline social activities, such as parties, hiking, sports activities, concerts, etc., and share information with other users.

Event-based social networks can provide users with online-to-offline social services, and help users initiate and develop personalized event participation plans. Users form online group relationships through common interests online, and initiate offline gathering events online. Event-based social networks have broader social attributes than location-based social networks. Existing work shows that event socialization in recommender systems The network has better recommendation features than traditional social networks.

Most of the current event-based social network recommendation is mainly based on the user's one-way perspective to extract feature preferences for recommendation, although the social influence of the event sponsor is considered, but the potential attractiveness of the event is insufficiently represented. On the other hand, regarding the influence of topic factors, only the event topic is used as one of the recommendation factors, and the user topic factor and its matching degree with the event topic are less considered.

SUMMARY OF THE INVENTION

In view of this, it is necessary to provide a personalized event recommendation method that combines several types of main context information to calculate user preferences and event potential preferences, and finally fuses topic matching and user-event bidirectional preferences. system.

A personalized event recommendation method integrating topic matching and bidirectional preference, comprising the following steps:

Step 1: Extract the topic information of the event with the document topic generation model LDA, and obtain the user topic information according to the historical event records that the user participated in, calculate the topic of the new event and the user's historical event, and use the cosine similarity to calculate the topic matching of the user-event pair. degree score;

Step 2: Build a user preference model and an event preference model respectively, and calculate the user preference score and the event preference score respectively;

Step 3: Use the Bayesian personalized ranking algorithm BPR to learn the weight parameters of the user preference score and the event preference score, obtain the user event bidirectional preference score, and obtain the user-event pair by linearly weighted combination of the topic matching score and bidirectional preference score. The final recommendation score is to recommend the top K events after sorting to the user.

Further, the document topic generation model LDA in step 1 has a three-layer generative Bayesian network structure, including documents, topics and words, wherein document-topic and topic-word are subject to polynomial distribution; Probabilistically select a topic, and select a word from this topic with a certain probability. The topics in any document conform to the Dirichlet distribution, and the relationship between texts is explored through this distribution.

Further, in the calculation of the subject of the new event and the user's historical event described in step 1, the cosine similarity is used to calculate the subject matching score of the user-event pair, and the specific steps include:

Step 1-1, all event description contents are formed into a document set D and the stop words are removed, the document set D is input into the document topic generation model LDA, and the topic distribution of each event is obtained respectively;

Remove stop words and punctuation marks from all event content, treat the document content after removing noise interference words as the set D of all documents, input it into the LDA topic model, and generate the joint distribution p(ω of the topic and words of the document d _i ) ,z|α,β), as shown in formula (1):

和主题词分布ν；Two unknown parameters in the model are then estimated using the Gibbs sampling method: the event topic distribution

and subject heading distribution ν;

Step 1-2, calculate the similarity of topic distribution between the target user's historical events and new events according to the JS divergence algorithm;

给定事件e_dp和e_dq分别具有主题分布

通过JS散度方法首先计算两者之间的JS散度

如式(2)所示：The topic distribution of all events has been generated according to equation (1)

Given events e _dp and e _dq have topic distributions respectively

First calculate the JS divergence between the two by the JS divergence method

As shown in formula (2):

Among them, D _js ∈ [0,1], D _KL represents the KL divergence, which is used to describe the difference between the two probability distributions p and q. The calculation formula is shown in formula (3):

Combining equations (2) and (3), the topic similarity of events _{edp and edq can} be obtained as S _topic , as shown in equation (4):

Among them, the value of the topic similarity S _topic of the event is located in [0, 1], and the closer the value is to 1, the higher the event similarity;

Steps 1-3, take the average of the similarity of all historical events of the target user, and obtain the theme matching score between the user and the new event;

作为用户和新事件的主题匹配度评分,如式(5)所示：The number of historical events of the target user is represented by E _u , and the average value of all the similarities of the target user is taken.

As the topic matching score between users and new events, as shown in formula (5):

来度量目标用户与新事件之间的主题匹配关系。According to the constructed topic matching model, finally

to measure the topic matching relationship between target users and new events.

Further, the construction of the user preference model in step 2 constructs the user's single-factor preference from three aspects: geographic location, social relationship, and time factors, specifically including:

Step 2-1-1, build a geographic location preference model:

如式(6)所示：The geographic location preference model calculates the probability that the target user will participate in the event held at the location, uses the kernel density estimation KDE method to model the two-dimensional geographic location distribution of the events that the user participates in, and uses the normalized event participation probability to represent the user's Geographical preference. The latitude and longitude coordinates of the event location are represented by (Lx,Ly), and the set of locations where the user has participated in the event in the past is represented by L(u), then the KDE function of user u

As shown in formula (6):

Among them, l _i =(Lx _i ,Ly _i ) ^T represents the two-dimensional vector of the latitude and longitude coordinates of the event location, m _l (u, l _i ) represents the frequency of the user u participating in the activities held at the geographic location _li , and σ represents the neighborhood The size of the window (bandwidth), N represents the number of position samples, and K( ) represents the Gaussian kernel function, which is defined in the form of equation (7):

Combining equations (6) and (7), the probability of user u participating in the event to be held at location l can be defined, as shown in equation (8):

The probability is normalized to obtain the user's preference score S _G (u, l) about the geographic location, as shown in Equation (9).

Among them, the denominator represents the maximum event participation probability of the target user;

Step 2-1-2, build a social relationship preference model:

In the user's social relationship network, the user will join at least one or more interest groups online, and choose to participate in events published by different groups, and judge the user's social relationship preference through the user's online same-group relationship. The group relationship mainly includes two kinds of interaction relationships;

可表示成式(10)所示：The first is the correlation between users and groups, which is defined as the interaction between users and all groups they belong to and between users and events created within the group, with G(u) representing the group to which the event that user u participates in belongs to , the user-group correlation

It can be expressed as formula (10):

Among them, m _p (u, g) represents the set of events and activities that the user u has participated in in the user group;

The second is the intra-group user correlation. The intra-group user correlation is defined by the similarity of friends in the target user’s group, and the similarity s(u, g) between the target user and the group is calculated, as shown in formula (11) shown:

Among them, sim(u _i , u _j ) represents the similarity between user u _i and user u _j in the same group, as shown in formula (12);

如式(13)所示：Normalize s(u,g) to

As shown in formula (13):

Combining the above two interaction relationships, users who belong to the same group tend to participate in events created by other users in these groups. Combining the correlation between users and groups and the correlation between users in the group, the social interaction of user u about online group g is obtained. The preference score S _I (u, g) is shown in formula (14):

Among them, α∈[0,1] is used as the weight control parameter. In the social relationship network, it is equally important to set the preference association between the target user and the group and the association between the users in the group. The value of α here is set as is 0.5;

Step 2-1-3, build the time factor preference model:

当新事件在一周的某个特定时间段中发生时，则将该时间段的向量分量值置为1，否则为0；在时间偏好模型中根据用户参加的历史事件记录将用户表示为用户时间向量

如式(15)所示：The time factor of the event is an important preference factor that needs to be considered when calculating user preferences; the new event e that the user can choose to participate in is represented as a 7*24-dimensional event time vector

When a new event occurs in a specific time period of the week, the vector component value of the time period is set to 1, otherwise it is 0; in the time preference model, the user is represented as the user time according to the historical event records that the user participated in vector

As shown in formula (15):

Among them, E _u represents the set of historical events that the target user has participated in, and then calculate the cosine similarity s(u, e) between the user time vector and the new event time vector, as shown in Equation (16):

For a new event e, user u _i ∈ U can obtain the similarity s(u _i ,e) according to formula (16), and normalize the similarity to obtain the user’s time preference score for the event S _T (u _i ,e ), as shown in formula (17):

Further, the calculating user preference score in step 2 specifically includes:

For the geographic location preference model, the geographic location preference score is expressed by predicting the probability of the user participating in the event held at the location; for the social relationship preference model, the relationship between the target user and the group and the correlation with the users in the group are two factors. Calculate the social preference score of the target user; for the time factor preference model, a unified vector representation with two granularities of date and hour is constructed, and based on this, the similarity of the user-event pair is calculated as the time preference score of the target user; combined with These three single-factor preferences form a user preference perception model, and linearly combine the three single-factor preferences to obtain the overall preference score S _user of user u to event e, as shown in formula (18):

Among them, _SG , _SI , and _ST represent the user's preference score on the three single factors of geographic location, social relationship, and time factor, respectively.

Further, the construction of the event preference model in step 2 constructs the single-factor preference of the event from two aspects, the popularity of the event location and the influence of the event sponsor, specifically including:

Step 2-2-1, build an event location popularity preference model:

Calculate the popularity of geographic location according to user u and users in the online group g he joined to visit the place;

First, define the popularity p( _le , u) of the event geographic location _le with respect to the user u, as shown in Equation (19):

Among them, the numerator m _l (u, _{le ) is the frequency of the user u participating in the activities held at the geographical location 1 e ,} and the denominator is the maximum frequency of the location that the user u has visited in the past; similarly, the geographical location 1 _e is defined as to where the user u is located. The popularity p( _le ,g) of group g is shown in formula (20):

Among them, the numerator represents the frequency of each user in group g participating in practical activities at location l, and the denominator is the maximum frequency of locations visited by group members in the past, from which the popularity of geographic location _le with respect to users in group g can be calculated. ; Combining p( _le ,u) and p( _le ,g) to define the total popularity of the location of the event to be recommended to the target user u as P( _le ,u,g), as shown in formula (21) :

P(le ,u,g)= _αp ( _le ,u)+(1-α)p( _le ,g) (21)

Step 2-2-2, construct the influence preference model of the event organizer:

First, the influence degree of the event sponsor on the target user, the implicit preference of the event is expressed by the reputation or influence degree of the sponsor; the influence degree I(e, u) of the event on the user u is defined, as shown in Equation (22) ) as shown:

Among them, m _h (u, u _h ) represents the set of events organized by the organizer u _h that the user u has participated in, and E _h is the set of all events organized by the organizer u _h ;

Second, the influence degree of the event organizer in the group, for the online group where the target user is located, the influence degree of the event in the group is represented by the proportion of the frequency of users participating in the group, and the influence degree of the user in the group is expressed as I( e, g) are represented, as shown in formula (23):

Among them, U _g represents the set of users in the group u _h , m _h (u _i , u _h ) represents the set of events organized by the organizer u _h that the user u _i participates in, and E _h (g) represents that u _h is in the group u _h The set of events held in the event organizer; the comprehensive influence score I(e, u, g) of the event organizer is obtained by combining the influence of the event organizer on the target users and users in the group, as shown in formula (24):

I(e,u,g)=αI(e,u)+(1-α)I(e,g) (24)

Further, the calculating event preference score in step 2 specifically includes:

For new events that have not occurred, the preference of events is expressed by calculating the event location popularity and event sponsor influence of the new event; for the constructed event location popularity P( _le ,u,g) and event sponsor influence Force I(e, u, g) is linearly combined, and the preference score S _events of event e to user u is calculated, as shown in formula (25):

Further, in step 3, the user event bidirectional preference score is obtained, and the linear weighted combination of the theme matching degree score and the bidirectional preference score is used to obtain the final recommendation score of the user-event pair, and the specific steps include:

Step 3-1, find bidirectional preference for user-event pair:

Assuming that the preference score weights of users and events are θ ₁ and θ ₂ respectively, the two-way preference score of user events is obtained by weighted fusion of the two _{, Su,e} = θ ₁ S _user + θ ₂ S _events ; the problem of bidirectional preference score is converted into Find the weight vector of the two preference scores, and choose to use the implicit feedback as the training data to learn the weight vector;

Select the learning algorithm BPR based on Bayesian maximum likelihood estimation to sort and learn the weights, and learn the correct sorting order of user-event pairs according to the user's implicit feedback data on events, so that the events that users participate in are ranked in new events or other events. Before; first, define the maximum posterior probability p(θ|R), as shown in Eq. (26):

p(θ|R)∝p(R|θ)p(θ) (26)

Among them, θ represents the weight vector, R represents the set of all user-event pairs, and p(R|θ) is defined as formula (27);

where R _u represents the user-event pair of user u, and p(e _i >e _j ) represents the probability that event e _i ranks ahead of e _j for user u, as shown in Equation (28):

p(e _i >e _j |θ)=σ(s(u,e _i )-s(u,e _j )) (28)

为了更方便进行优化，假设θ服从均值为0的正态分布，展开推导得出最终优化目标函数lnp(θ|R)，如式(29)所示：Among them, s(u,e) is the bidirectional preference score S _u,e ,

In order to make the optimization more convenient, assuming that θ obeys a normal distribution with a mean of 0, the final optimization objective function lnp(θ|R) can be derived from the expansion, as shown in Equation (29):

Among them, λ represents the regular term coefficient. The objective function is maximized through the implicit interactive feedback data of user events, and the optimal weight parameter vector is obtained. The stochastic gradient descent algorithm SGD is used to solve the optimization problem. The user-event pair of the target user is extracted to update the weight vector θ, and the update process is shown in formula (30):

Among them, α is the learning rate, s _ij =s(u,e _i )-s(u,e _j ); through the above learning process, the weight vector θ can be automatically obtained according to the user event preference score training set and hyperparameters α and λ , so as to obtain the bidirectional preference score S _u,e ;

Step 3-2, combine topic matching and bidirectional preference to obtain the final recommendation score of the user-event pair:

与用户事件双向偏好评分S_u,e线性加权求和得到最终的用户-事件对推荐度评分S_Rec，如式(31)所示：First, extract the event topic through the LDA topic model and obtain the topic matching score between the user and the event; secondly, according to the user event context information in the EBSN, the user and event preference models are respectively constructed, and the user event bidirectional preference score is obtained through the BPR learning algorithm. ; Finally, score the topic match

The final user-event pair recommendation score S _Rec is obtained by linearly weighted summation with the user event bidirectional preference score S _u,e , as shown in formula (31):

Among them, γ is a weight parameter, which is usually set manually based on experience, and the optimal setting will be determined through experiments.

And, an implementation system for personalized event recommendation that integrates theme matching and bidirectional preference, which is used to implement the personalized event recommendation method that integrates theme matching and bidirectional preference as described in any of the above, and the implementation system includes:

The document topic generation module is used to extract the topics of the user's historical events and new events, and calculate the topic distribution and word distribution of the events. One of the key factors is incorporated into the recommendation model for event recommendation;

Build a user preference module, which is used to construct the user's single-factor preference from three aspects of geographic location, social relationship, and time factors, and weighted and fused the three single-factor preferences to obtain the user's overall preference;

Build an event preference module, which uses the social influence of the event organizer in the group and the popularity of the geographical location where the event is held to represent the preference of the event;

The user event bidirectional preference scoring module uses the sorting learning algorithm to solve the weight parameters of the user preference score and the event preference score, and obtains the user event bidirectional preference score;

The final recommendation scoring module for user-event pairs is used to linearly weight the topic matching score and the bidirectional preference score to obtain the final recommendation score for the user-event pair.

Further, the user preference module includes a geographic location preference module, a social relationship preference module and a time factor preference module, and the event preference module includes an event location popularity preference module and an event sponsor influence preference module, wherein:

The geographic location preference module is used to represent the geographic location preference score by predicting the probability that the user participates in an event activity held in a certain geographic location;

The social relationship preference module is used to calculate the social preference score of the target user from two aspects: the relationship between the target user and the group and the correlation with the users in the group;

The time factor preference module is used to construct a unified vector representation of two granularities of date and hour, and calculate the similarity of the user-event pair as the time preference score of the target user;

The event location popularity preference module is used for when a new event is recommended, the location of the event is an important selection basis for interested users, which is called the popularity of geographic location in the user group, considering the influence of event geographic location. Popularity can more accurately calculate the attractiveness of events to users;

The event sponsor influence preference module is used to improve the accuracy of recommendation according to the influence of the event sponsor in the group where the target user is located, from the influence of the event sponsor on the target user and the event sponsor in the group. The influence is calculated from two aspects.

In the above-mentioned personalized event recommendation method and system based on fusion topic matching and bidirectional preference, first, the topic information of the event is extracted by using the document topic generation model LDA, and the user topic information is obtained according to the historical event records participated by the user, and the relationship between the user and the event is calculated. The topic matching degree is an important recommendation factor in the recommendation model, and the topic factor can better represent the feature preference; secondly, for the event-based social network recommendation, considering the bidirectional perspective of the user and the event, the preference model of the user and the event is constructed, respectively. The user preference score and the event preference score are used to mine the preference relationship more completely from the perspectives of users and events. Finally, the user-event pair matching degree is combined with the user-event bidirectional preference linear weighted combination to obtain the final user-event pair comprehensive score. The sorted top K (ie, TOP-K) user-event pairs are used as recommendation results. This scheme has carried out a large number of experiments on the Meetup real data set, and compared it with other event recommendation algorithms. So as to achieve the purpose of personalized recommendation.

Description of drawings

FIG. 1 is a structural diagram of an overall recommendation fusion framework of a method and system for personalized event recommendation that integrates topic matching and bidirectional preference according to an embodiment of the present invention.

FIG. 2 is a structural block diagram of a document topic generation model LDA of the method and system for personalized event recommendation that integrates topic matching and bidirectional preference according to an embodiment of the present invention.

Detailed ways

In this embodiment, a method for recommending personalized events that integrates topic matching and bidirectional preference is taken as an example, and the present invention will be described in detail below with reference to specific embodiments and accompanying drawings.

Please refer to FIG. 1 and FIG. 2 , which illustrate a method and system for recommending a personalized event that integrates topic matching and bidirectional preference provided by an embodiment of the present invention.

Here is a detailed explanation of the technical details involved in the personalized event recommendation system of the software's fusion theme matching and bidirectional preference. The main idea is that, first, the topics of new events and user historical events are calculated through the LDA topic model, and the topic matching degree of user-event pairs is calculated by using cosine similarity, and the user preference model and event preference model are respectively constructed. Among them, the user preference model calculates the user's comprehensive preference score from three aspects: time, geography, and social relationship. The event preference model expresses an event potential preference score based on the geographic popularity of the new event in the target user group and the sponsor's in-group social influence. Then, the Bayesian Personalized Ranking (BPR) algorithm is used to learn the weight parameters of the user preference score and the event preference score, and the user event bidirectional preference score is obtained. Finally, the linear weighted fusion with the topic matching degree is used to obtain the final recommendation score of the user-event pair, and the sorted TOP-K events are recommended to the user. That is, the software matches users and event topics, calculates user preferences and event potential preferences in combination with several types of main context information, and finally integrates topic matching and user-event bidirectional preferences for event recommendation.

1. A recommendation framework that integrates LDA topic matching and user event bidirectional preference

On the basis of the existing work, based on the geographic location information, time information, social relationship and other related user event context information in EBSN, an event combining user-event pair topic matching and user-event pair bidirectional preference is proposed. Recommended plan. In this scheme, the influence of topic matching degree, user preference and event preference on event recommendation is considered separately, and these factors are integrated to effectively recommend interest events to users. The overall framework of the recommendation model is shown in Figure 1, and the specific recommendation process is as follows:

1) According to the description document of the event in the EBSN, the LDA topic model is used to calculate the historical event topic of the new event and the target user, and the topic of the user's historical event is used to represent the user's topic, and then the semantic similarity between the event and the user topic distribution is calculated to obtain the user- Match score for the event topic.

2) Calculate the user preference score and the event preference score. For the user preference, the preference score is calculated and linearly integrated from the three aspects of geographic location, social relationship and time, and the event preference is based on the popularity of the event location and the event organizer. The social influence is represented by the same linear fusion to get the event preference score. It should be noted that when calculating the geographic location popularity and sponsor influence of an event, only the target user's group and users in the group are used, and the associations of other users and groups are ignored, so as to improve the recommendation performance and reduce the computational complexity. Spend.

3) Obtain the user-event theme matching score, the user's preference score for the event, and the event's preference score for the user through the above calculation. Firstly, the Bayesian personalized sorting algorithm is used to learn the weights of user preference scores and event preference scores, and then the bidirectional preference scores are obtained by integrating the preference scores of users and events according to the weights. Finally, the topic matching score and bidirectional preference score information are linearly combined to obtain the final User-event pair recommendation scores, and recommend the highest-rated TOP-K events to users.

2. LDA-based topic matching model

In the event social network, there is an obvious topic semantic similarity relationship between users and events. Users usually choose to participate in a certain type of interesting events. Generally, this type of events has similar attributes and themes. Applying the topic of events in the recommendation can better capture the preferences of users and events. We represent the topic of the user with the topic of the historical events that the user participated in, and calculate the topic distribution and word distribution of new events. The topic similarity of is the topic matching degree, which is incorporated into the recommendation model as one of the key factors of recommendation for event recommendation.

When two documents have the same features such as topics, it is difficult to distinguish these two objects with the TF-IDF (Term Frequency–Inverse Document Frequency) algorithm, so a Bayesian-based LDA topic model is selected to calculate the document topic distribution and word distribution . The LDA topic model is a Bayesian probabilistic model for computing document topic distributions for clustering latent topics for documents and generating document topics. The core idea is that each document selects a certain topic with a certain probability, and selects a certain word from this topic with a certain probability. It is believed that the topics in any document conform to the Dirichlet distribution, and through this distribution, we can discover the difference between texts. Relationship. LDA consists of a three-layer generative Bayesian network structure, including documents, topics, and words, and both document-topic and topic-words obey multinomial distributions. The LDA topic model generation process is shown in Figure 2.

再从主题z所对应的词分布

中抽取一个单词 w，重复以上过程N_m次即可生成文档d_i。在这个过程中利用Gibbs采样方法即可求解文档d_i的主题分布。Given a set of documents D={d ₁ , d ₂ ,...,d _m }, v and Represents the topic distribution of the document d _i and the prior Dirichlet distribution of the word distribution, respectively, α, β are the hyperparameters of the topic prior distribution and word prior distribution given by experience, respectively, k is the number of topics in the document set specified in advance , N _m represents the total number of words in document d _i , and M is the number of documents in the document set. For each word in the document d _i , LDA determines the topic distribution v of the document according to the prior knowledge α, then extracts a topic z from the topic distribution v, and determines the word distribution of the current topic according to the prior knowledge β

Then from the word distribution corresponding to topic z

Extract a word w from , and repeat the above process N _m times to generate a document d _i . In this process, the topic distribution of the document _di can be solved by using the Gibbs sampling method.

和主题词分布v。To calculate the topic similarity between users and events according to LDA, first convert the text content into semantic features, and use the LDA topic model to calculate the topic distribution for each event. The content of the event is mainly composed of title and description documents, and also includes information such as time and venue. The theme of the event can be extracted from the content of the event. In contrast, users can choose to set interest tags to express their preferences. However, many users do not set interest tags or self-introduction content. User content lacks document information and faces extremely sparse data. At this time, there are no available features to represent user topics. Therefore, it is more accurate to choose the theme of the historical events that the user participated in to express the user theme, and avoid the problems of sparse data and blank labels. Remove stop words and punctuation marks from all event content, regard the document content after removing noise interference words as the set D of all documents, input it into the LDA topic model, and generate the topics and words of document d _i according to the generation process described above The joint distribution p(ω,z|α,β) of , as shown in formula (1). The Gibbs sampling method is then used to estimate two unknown parameters in the model, the event topic distribution

and subject heading distribution v.

给定事件e_dp和e_dq分别具有主题分布

通过JS散度方法首先计算两者之间的JS散度

如式(2)所示。After obtaining the topic distribution and word distribution of the event document through the LDA process, the JS divergence (JensenShannon divergence) method is used to calculate the similarity between events according to the topic distribution of the event. JS divergence is a variant based on KL divergence (Kullback-Leibler divergence), which is symmetric and solves the asymmetric problem of KL divergence, which can better measure the similarity of two probability distributions. The topic distribution of all events has been generated according to equation (1)

Given events e _dp and e _dq have topic distributions respectively

First calculate the JS divergence between the two by the JS divergence method

As shown in formula (2).

Among them, D _js ∈ [0,1], D _KL represents the KL divergence, which is used to describe the difference between the two probability distributions p and q. The calculation formula is shown in formula (3).

Combining Equation (2) and Equation (3), the topic similarity of events _{edp and edq can} be obtained as S _topic , as shown in Equation (4).

作为用户和新事件的主题匹配度评分,如式(5)所示。The value of the topic similarity S _topic of the event is in [0, 1], and the closer the value is to 1, the higher the event similarity. As mentioned earlier, the topic similarity between new events and user historical events is regarded as the topic similarity between users and events, and users often participate in multiple events, and there are multiple topic similarities between new events and E _u Represents the number of historical events of the target user, and takes the average of all the similarities of the target user

As the topic matching score between users and new events, as shown in Equation (5).

表示主题的单词分布，

表示文档主题分布，Dir()表示Dirichlet分布，Mult()表示多项式分布，Poiss() 表示泊松分布。Algorithm 1 describes the process of calculating the topic matching degree of user-event pairs through the LDA topic model, where

the word distribution representing the topic,

Represents the document topic distribution, Dir() represents the Dirichlet distribution, Mult() represents the multinomial distribution, and Poiss() represents the Poisson distribution.

Algorithm 1 presents the process of using LDA topic model and JS divergence algorithm to solve the score of user-event to topic matching degree. First, form a document set with all event descriptions and remove stop words as the input of the LDA model to obtain the topic distribution of each event (lines 2 to 11); then calculate the target user according to the JS divergence algorithm The topic distribution similarity between historical events and new events (Lines 12 to 14); finally, the similarity of all historical events of the target user is averaged to obtain the topic matching score between the user and the new event (Line 15) to line 16).

3. User-based preference model

For user preference, feature learning is generally performed from the relevant context information of the user, and the learned feature information is expressed as user preference. The following is to construct the user's single-factor preference from the three aspects of geographical factors, social relations, and time factors, and the three single-factor preferences are weighted and integrated to obtain the user's overall preference.

3.1 Geographical Preferences

如式(6) 所示。The geographic location preference model calculates the probability that the target user will participate in the event at this location, and uses the KDE (Kernel Density Estimation, Kernel Density Estimation) method to model the two-dimensional geographic location distribution of the events that the user participates in. The event participation probability represents the user's preference for geographic location. The latitude and longitude coordinates of the event location are represented by (Lx, Ly), and the set of locations where the user has participated in the event in the past is represented by L(u), then the KDE function of user u

As shown in formula (6).

Among them, l _i =(Lx _i ,Ly _i ) ^T represents the two-dimensional vector of the latitude and longitude coordinates of the event location, m _l (u, l _i ) represents the frequency of the user u participating in the activities held at the geographic location _li , and σ represents the neighborhood The size of the window (bandwidth), N represents the number of position samples, K(·) represents the Gaussian kernel function, and its definition form is shown in formula (7).

Combining Equation (6) and Equation (7), the probability of user u participating in the event to be held at location l can be defined, as shown in Equation (8).

The probability is normalized to obtain the user's preference score S _G (u, l) about the geographic location, as shown in Equation (9).

3.2 Social relationship preferences

In the user's social relationship network, the user generally joins at least one or more interest groups online, and can choose to participate in events published by different groups. In these group relationships, users usually choose the preference group they are most interested in to participate in, and members in the same group generally have the same interests. Therefore, users can be considered through the user’s online same-group relationship. The preference of social relations mainly includes two kinds of interaction relations.

可表示成式(10)所示。1) Relevance of users and groups. That is, the interaction between the user and all the groups they belong to, and between the user and the events created within the group. Let G(u) denote the set of groups to which the events that user u participates in belong to, then the correlation between users and groups

It can be expressed as formula (10).

Among them, m _p (u, g) represents the set of events and activities that the user u has participated in in the user group.

2) User relevance within the group. The intra-group user correlation is defined by the similarity of friends in the target user's group, and the similarity s(u, g) between the target user and the users in the group is calculated, as shown in Equation (11).

Among them, sim(u _i , u _j ) represents the similarity between user u _i and user u _j in the same group, as shown in equation (12).

如式(13)所示。Finally, s(u,g) is normalized to

As shown in formula (13).

Combining these two interaction relationships, users belonging to the same or similar groups tend to participate in events created by these groups, and the social preferences of user u about online group g are obtained by combining user-group correlations and intra-group user correlations The score S _I (u, g) is shown in formula (14).

Among them, α∈[0,1] is used as the weight control parameter. In the social relationship network, it is generally considered that the preference association between the target user and the group is equally important as the association between the users in the group. Through experimental verification, the value of α here is set as is 0.5.

3.3 Time preference

The time factor of events is another important preference factor to consider when calculating user preferences. Different users have different preferences when choosing to participate in events. Some users may prefer to participate in activities in the evening, while others may prefer to participate in activities in the morning, or prefer different time points on weekdays or weekends. In reality, time is cyclical, mainly 7 days a week and 24 hours a day. For users who choose to participate in activities on a certain day of the week and a few hours in a day, two different granularity levels will be formed. user time preference. We represent users' temporal preferences by combining user choices at two levels of granularity.

当新事件在一周的某个特定时间段中发生时，即将该时间段的向量分量值置为1，否则为0。因此，可以在时间偏好模型中根据用户参加的历史事件记录将用户表示为用户时间向量

如式(15)所示。If a user chooses a certain time period on a certain day of the week to participate in the event, this may represent an implicit time preference of the user, and the user may choose to participate in the event again at the same time period next time. In order to express this implicit preference uniformly and intuitively, we represent the new event e that the user can choose to attend as a 7*24-dimensional event time vector

When a new event occurs in a certain time period of the week, the value of the vector component of the time period is set to 1, otherwise it is 0. Therefore, the user can be represented as a user time vector in the time preference model according to the historical event records that the user participated in

As shown in formula (15).

Among them, E _u represents the set of historical events that the target user has participated in, and then calculates the cosine similarity s(u, e) between the user time vector and the new event time vector, as shown in Equation (16).

For a new event e, user u _i ∈ U can obtain the similarity s(u _i ,e) according to formula (16), and normalize the similarity to obtain the user’s time preference score for the event S _T (u _i ,e ), as shown in formula (17).

3.4 User Fusion Preference Score

According to the previous modeling of the user's single-factor preference model from three aspects, the user's preference scores on geographic location, social relationship and time are calculated respectively. For geographic location, the geographic location preference score is expressed by predicting the probability of users participating in events held at the location; for social relations, the social preferences of target users are calculated from the relationship between the target user and the group and the correlation with users in the group. Scoring; for time preference, a unified vector representation with two granularities of date and hour is constructed, and the similarity of user-event pairs is calculated based on this as the target user's time preference score. Combine these three single-factor preferences to form a user preference perception model, and linearly combine the three single-factor preferences to obtain the overall preference score S _user of user u to event e, as shown in formula (18).

Among them, the distribution of _SG , _SI , and _ST represents the user's preference score on the three factors of geographic location, social relationship, and time. Algorithm 2 describes the calculation process of user preference score.

Algorithm 2 presents the process of calculating the user's comprehensive preference score by combining the user's preference in three factors: geographic location, social relationship, and time. The probability that a user may participate in an event held at a specific location is predicted by the kernel density estimation algorithm, and the probability is normalized to represent the user's geographic preference (line 3). The social relatedness of the upper group and members in the group represents the social preference (Lines 5 to 11); the new events and the user’s historical events are represented as time vectors, and the cosine similarity between the two is calculated to represent the user’s time preference (Line 4). row); finally, the user's total preference score is obtained by linearly combining the three preference values (rows 13 to 14).

4. Event-Based Preference Models

For event preferences, consider learning from event sponsors and event ontology information. Since events lack active personalized contextual information compared to users, for new events, there is no information such as history records, personalized tags, etc. Therefore, the social influence of the event organizer in the group and the geographic location of the event The popularity of the location in the group to represent the preference of the event.

4.1 Event Location Popularity

The geographic location of the event is a consideration in the user's choice of whether to participate in the event. An online group that a user joins is generally a group of users with the same interests, and there may be multiple users who choose to participate in the same event. Therefore, for the recommendation of a new event, the venue of the event can be important to interested users. The selection basis of , and this relationship is called the popularity of geographic location in the user group. Considering the popularity of an event's geographic location in a model for calculating event preferences can more accurately calculate the attractiveness of an event to users. The popularity of the geographic location is calculated according to the frequency of location visits by user u and users in the online group g he joins.

First, define the popularity p( _le , u) of the event geographic location _le with respect to the user u, as shown in Equation (19).

Among them, the _numerator m _l (u, _le ) is the frequency of the user u participating in the activities held at the geographic location le, and the denominator is the maximum frequency of the location that the user u has visited historically. Similarly, the popularity p( _le ,g) of the geographic location _le with respect to the group g where the user u belongs can be defined, as shown in equation (20).

Among them, the numerator represents the frequency of each user in group g participating in practical activities at location l, and the denominator is the maximum frequency of locations visited by group members in the past, from which the popularity of geographic location _le with respect to users in group g can be calculated. . Combining p( _le , u) and p( _le , g), the total popularity of the location of the event to be recommended to the target user u can be defined as P( _le , u, g), as shown in Equation (21) .

P(le ,u,g)= _αp ( _le ,u)+(1-α)p( _le ,g) (21)

4.2 Influence of event organizers

In the event social network, the initiator of each event is also an ordinary user on the network. Generally, the organizer initiates a certain event and gets a good response. Then the next time another new event is launched, the users who participated in the previous event are very likely to Choose to revisit its events. Although the event to be recommended is a brand-new event that has not yet happened for each user, the organizer of the event may be an active organizer of this type of event, and may have hosted many events before, which is very important for solving the problem of event recommendation. The cold start problem provides more auxiliary recommendation information. It can be seen that the influence of the event organizer among the user groups in the group is an important feature of the event preference. This software improves the accuracy of recommendation according to the influence of the event organizer in the target user group. Its influence can be considered from the following two aspects.

1) The influence of the event organizer on the target users. In the event social network, there is no user's rating information for the event, and the influence of the organizer and the event cannot be intuitively represented, and it is meaningless to score the event at the end of its life cycle, because it will not affect subsequent events. new event, so choose to express the implicit preference of the event through the sponsor's reputation or influence. First, define the influence degree I(e, u) of an event on user u, as shown in Equation (22).

Among them, m _h (u, u _h ) represents the set of events organized by the organizer u _h that the user u has participated in, and E _h is the set of all events organized by the organizer u _h .

2) The influence of the event organizer in the group. For the online group where the target user belongs, the influence degree of the event in the group can be similarly expressed by the frequency ratio of the user’s participation, and the influence degree of the user in the group is expressed by I(e, g), as shown in Equation (23) shown.

Among them, U _g represents the set of users in the group u _h , m _h (u _i , u _h ) represents the set of events organized by the organizer u _h that the user u _i participates in, and E _h (g) represents that u _h is in the group u _h A collection of events held in . Combined with the influence of the event sponsor on target users and users in the group, the comprehensive influence score I(e, u, g) of the event sponsor can be obtained, as shown in formula (24).

I(e,u,g)=αI(e,u)+(1-α)I(e,g) (24)

4.3 Event latent preference score

For new events that have not occurred, the software sets two key factors to attract users to participate in the geographical location and the influence of the organizer. Event preferences are represented by calculating the geographic popularity of a new event and the social influence of its sponsors. In order to reduce the computational complexity and avoid the interference and influence of weakly correlated data, the geographical location popularity of the event and the social influence of the sponsor are only limited to the group where the target user is located. It is assumed here that the remaining users or groups have zero relevance and have no effect on event preference. A linear combination of the event geographic location popularity P(le, u, g) and the sponsor's influence I( _e , u, g) constructed above is used to obtain the preference score S _events of event e to user u, as shown in the formula ( 25) shown.

Algorithm 3 details the process of calculating an event's latent preference score from the event location popularity and sponsor influence.

Algorithm 3 presents the process of solving the potential preference score of the event according to the geographic location popularity of the event and the influence of the sponsor. For the group where the target user belongs, calculate the popularity of the event geographic location for the user and the group according to Equation (19) and Equation (20), respectively, and combine the two to represent the total popularity of the event geographic location (rows 3 to 8); Similarly, the influence of the event organizer on users and groups is obtained from equations (22) and (23) (rows 9 to 13), and the two are combined to express the influence of the event organizer; finally, the location popularity and The linear combination of sponsor influence yields the event's latent preference score (Line 17).

5. A recommendation algorithm that integrates topic matching and user event bidirectional preference

We have used the LDA topic model to solve the topic distribution of users and events respectively, and calculated the topic matching degree of the user-event pair according to the topic distribution. Next, we built a feature preference scoring model for users and events, and obtained the user preferences respectively. Scoring and Event Preference Scoring. Now the topic matching and user event preference are fused to obtain the final recommendation score. The first step is to use the ranking learning algorithm to solve the weight parameters of the user preference score and the event preference score to obtain the user event bidirectional preference score; The linearly weighted combination of topic matching score and bidirectional preference score yields the final recommendation score for the user-event pair. The following is a specific introduction.

1) Obtain bidirectional preference scores for user-event pairs. Assuming that the preference score weights of users and events are θ ₁ and θ ₂ respectively, the two-way preference scores of user events are obtained by weighted fusion of the two _{, Su,e} = θ ₁ S _user + θ ₂ S _events . Therefore, the key problem of the bidirectional preference score is to find the weight vector of the two preference scores, and choose to use the implicit feedback as the training data to learn the weight vector. Unlike explicit feedback where users rate items, implicit feedback in event social networks can only be represented by the interaction information between the user and the event, that is, if the user participated in the event, the feedback is 1, otherwise the feedback is 0. Apparently, the user feedback is 0 for all new events.

Here, the learning algorithm BPR based on Bayesian maximum likelihood estimation is selected to sort and learn the weights, and learn the correct sorting order of user-event pairs according to the implicit feedback data of users to events, so that the events that users participate in are ranked in the new events or before other events. First, define the maximum posterior probability p(θ|R), as shown in Equation (26).

p(θ|R)∝p(R|θ)p(θ) (26)

Among them, θ represents the weight vector, R represents the set of all user-event pairs, and p(R|θ) is defined as shown in Equation (27).

Among them, R _u represents the user-event pair of user u, and p(ei > e _j ) represents the probability that event _ei ranks ahead of e _j for user _u , as shown in Equation (28).

p(e _i >e _j |θ)=σ(s(u,e _i )-s(u,e _j )) (28)

为了更方便进行优化，假设θ服从均值为0的正态分布，展开推导得出最终优化目标函数lnp(θ|R)，如式(29)所示。Among them, s(u,e) is the bidirectional preference score S _u,e ,

In order to make the optimization more convenient, assuming that θ obeys a normal distribution with a mean value of 0, the final optimization objective function lnp(θ|R) can be derived from the expansion, as shown in Equation (29).

Among them, λ represents the regularization term coefficient. The optimal weight parameter vector can be obtained by maximizing the optimization objective function through the implicit interactive feedback data of user events. Stochastic Gradient Descent (SGD) is used to solve the optimization problem. In the iterative process, the user-event pairs of the target user are randomly extracted from the training set to update the weight vector θ. The update process is shown in Equation (30).

where α is the learning rate, s _ij =s(u,e _i )-s(u,e _j ). Through the above learning process, the weight vector θ can be obtained automatically according to the user event preference score training set and hyperparameters α and λ, thereby obtaining the bidirectional preference score _Su,e .

与用户事件双向偏好评分S_u,e线性加权求和得到最终的用户-事件对推荐度评分S_Rec，如式(31)所示。2) Combining topic matching and bidirectional preference to obtain the final recommendation score of the user-event pair. Based on the above discussion on topic matching and preference calculation between users and events, first, the topic of events is extracted through the LDA topic model and the topic matching score of users and events is obtained; secondly, users and events are constructed according to the context information of user events in EBSN. The preference model of events, the bidirectional preference score of user events is obtained through the BPR learning algorithm; finally, the topic matching degree is scored

The final user-event pair recommendation score S _Rec is obtained by linearly weighted summation with user event bidirectional preference score S _u,e , as shown in formula (31).

Among them, γ is a weight parameter, which is usually set manually based on experience, and the optimal setting will be determined through experiments. Algorithm 4 describes the process of integrating topic matching and bidirectional preference to obtain the final recommendation score of user-event pairs.

Algorithm 4 presents the final fusion process of topic matching score and user event bidirectional preference score. First, the training set generated by the user preference score set and the event preference score set is sorted and learned by the Bayesian personalized ranking algorithm, and the optimal weight vector θ is obtained, and the bidirectional preference of the target user's user-event pair is calculated according to θ. score (lines 2 to 10); secondly, linearly combine the topic matching score and bidirectional preference score of the user-event pair to get the final recommendation score (line 11 to 13), so as to sort according to the final recommendation score Recommend TOP-K events to users.

So far, we have combined topic matching and user event bidirectional preference to propose a personalized event recommendation scheme, and its specific content is introduced in detail in the above section.

And, an implementation system for personalized event recommendation that integrates theme matching and bidirectional preference, which is used to implement the personalized event recommendation method that integrates theme matching and bidirectional preference as described in any of the above, and the implementation system includes:

The document topic generation module is used to extract the topics of the user's historical events and new events, and calculate the topic distribution and word distribution of the events. One of the key factors is incorporated into the recommendation model for event recommendation;

Build a user preference module, which is used to construct the user's single-factor preference from three aspects of geographic location, social relationship, and time factors, and weighted and fused the three single-factor preferences to obtain the user's overall preference;

Build an event preference module, which uses the social influence of the event organizer in the group and the popularity of the geographical location where the event is held to represent the preference of the event;

The user event bidirectional preference scoring module uses the sorting learning algorithm to solve the weight parameters of the user preference score and the event preference score, and obtains the user event bidirectional preference score;

The final recommendation scoring module for user-event pairs is used to linearly weight the topic matching score and the bidirectional preference score to obtain the final recommendation score for the user-event pair.

Further, the user preference module includes a geographic location preference module, a social relationship preference module and a time factor preference module, and the event preference module includes an event location popularity preference module and an event sponsor influence preference module, wherein:

The geographic location preference module is used to represent the geographic location preference score by predicting the probability that the user participates in an event activity held in a certain geographic location;

The social relationship preference module is used to calculate the social preference score of the target user from two aspects: the relationship between the target user and the group and the correlation with the users in the group;

The time factor preference module is used to construct a unified vector representation of two granularities of date and hour, and calculate the similarity of the user-event pair as the time preference score of the target user;

The event location popularity preference module is used for when a new event is recommended, the location of the event is an important selection basis for interested users, which is called the popularity of geographic location in the user group, considering the influence of event geographic location. Popularity can more accurately calculate the attractiveness of events to users;

The event sponsor influence preference module is used to improve the accuracy of recommendation according to the influence of the event sponsor in the group where the target user is located, from the influence of the event sponsor on the target user and the event sponsor in the group. The influence is calculated from two aspects.

In the above-mentioned personalized event recommendation method and system based on fusion topic matching and bidirectional preference, first, the topic information of the event is extracted by using the document topic generation model LDA, and the user topic information is obtained according to the historical event records participated by the user, and the relationship between the user and the event is calculated. The topic matching degree is an important recommendation factor in the recommendation model, and the topic factor can better represent the feature preference; secondly, for the event-based social network recommendation, considering the bidirectional perspective of the user and the event, the preference model of the user and the event is constructed, respectively. The user preference score and the event preference score are used to mine the preference relationship more completely from the perspectives of users and events. Finally, the user-event pair matching degree is combined with the user-event bidirectional preference linear weighted combination to obtain the final user-event pair comprehensive score. The sorted TOP-K user-event pairs are used as recommendation results. This scheme has carried out a large number of experiments on the Meetup real data set, and compared it with other event recommendation algorithms. So as to achieve the purpose of personalized recommendation.

It should be noted that the above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims (10)

1. a personalized event recommendation method of fusion theme matching and bidirectional preference, is characterized in that, comprises the following steps: Step 1: Extract the topic information of the event with the document topic generation model LDA, and obtain the user topic information according to the historical event records that the user participated in, calculate the topic of the new event and the user's historical event, and use the cosine similarity to calculate the topic matching of the user-event pair. degree score; Step 2: Build a user preference model and an event preference model respectively, and calculate the user preference score and the event preference score respectively; Step 3: Use the Bayesian personalized ranking algorithm BPR to learn the weight parameters of the user preference score and the event preference score, obtain the user event bidirectional preference score, and obtain the user-event pair by linearly weighted combination of the topic matching score and bidirectional preference score. The final recommendation score is to recommend the top K events after sorting to the user. 2. the personalized event recommendation method of fusion theme matching and bidirectional preference as claimed in claim 1, it is characterized in that, described document theme generation model LDA in step 1 has three-layer generative Bayesian network structure, including document , topic and word, where both document-topic and topic-word obey a multinomial distribution; each document selects a topic with a certain probability, and selects a word from this topic with a certain probability, and the topics in any document conform to the Dirichlet distribution , and explore the relationship between texts through this distribution. 3. the personalized event recommendation method of fusion theme matching and bidirectional preference as claimed in claim 2, is characterized in that, described in step 1 calculates the theme of new event and user historical event, adopts cosine similarity to calculate user-event To score the topic matching degree of the pair, the specific steps include: Step 1-1, all event description contents are formed into a document set D and the stop words are removed, the document set D is input into the document topic generation model LDA, and the topic distribution of each event is obtained respectively; Remove stop words and punctuation marks from all event content, treat the document content after removing noise interference words as the set D of all documents, input it into the LDA topic model, and generate the joint distribution p(ω) of the topic and words of the document d _i , z|α, β), as shown in formula (1):

Two unknown parameters in the model are then estimated using the Gibbs sampling method: the event topic distribution

and subject term distribution v; Step 1-2, calculate the similarity of topic distribution between the target user's historical events and new events according to the JS divergence algorithm; The topic distribution of all events has been generated according to equation (1)

Given events e _dp and e _dq have topic distributions respectively

First calculate the JS divergence between the two by the JS divergence method

As shown in formula (2):

Among them, D _js ∈ [0, 1], D _KL represents the KL divergence, which is used to describe the difference between the two probability distributions p and q. The calculation formula is shown in formula (3):

Combining equations (2) and (3), the topic similarity of events _{edp and edq can} be obtained as S _topic , as shown in equation (4):

Among them, the value of the topic similarity S _topic of the event is located in [0, 1], and the closer the value is to 1, the higher the event similarity; Steps 1-3, take the average of the similarity of all historical events of the target user, and obtain the theme matching score between the user and the new event; The number of historical events of the target user is represented by E _u , and the average value of all the similarities of the target user is taken.

As the topic matching score between users and new events, as shown in formula (5):

According to the constructed topic matching model, finally

to measure the topic matching relationship between target users and new events. 4. the personalized event recommendation method of fusion theme matching and bidirectional preference as claimed in claim 3, it is characterized in that, described in step 2 constructs user preference model from three aspects of geographical location, social relationship, time factor respectively. Build the user's single-factor preferences, including: Step 2-1-1, build a geographic location preference model: The geographic location preference model calculates the probability that the target user will participate in the event held at the location, uses the kernel density estimation KDE method to model the two-dimensional geographic location distribution of the events that the user participates in, and uses the normalized event participation probability to represent the user's Geographical preference. The latitude and longitude coordinates of the event location are represented by (Lx, Ly), and the set of locations where the user has participated in the event in the past is represented by L(u), then the KDE function of user u

As shown in formula (6):

Among them, l _i =(Lx _i , Ly _i ) ^T represents the two-dimensional vector of the latitude and longitude coordinates of the event location, m _l ( _u , li ) represents the frequency of user _u participating in activities held at the geographic location li , and σ represents the neighborhood The size of the window (bandwidth), N represents the number of position samples, and K( ) represents the Gaussian kernel function, which is defined in the form of equation (7):

Combining equations (6) and (7), the probability of user u participating in the event to be held at location l can be defined, as shown in equation (8):

The probability is normalized to obtain the user's preference score S _G (u, l) about the geographic location, as shown in formula (9).

Among them, the denominator represents the maximum event participation probability of the target user; Step 2-1-2, build a social relationship preference model: In the user's social relationship network, the user will join at least one or more interest groups online, and choose to participate in events published by different groups, and judge the user's social relationship preference through the user's online same-group relationship. The group relationship mainly includes two kinds of interaction relationships; The first is the correlation between users and groups, which is defined as the interaction between users and all groups they belong to and between users and events created within the group, with G(u) representing the group to which the event that user u participates in belongs to , the user-group correlation

It can be expressed as formula (10):

Among them, m _p (u, g) represents the event activity set that user u has participated in in the user group; The second is the intra-group user correlation. The intra-group user correlation is defined by the similarity of the friends in the target user’s group, and the similarity s(u, g) between the target user and the users in the group is calculated, as shown in formula (11) shown:

Among them, sim(u _i , u _j ) represents the similarity between user u _i and user u _j in the same group, as shown in formula (12);

Normalize s(u, g) to

As shown in formula (13):

Combining the above two interaction relationships, users who belong to the same group tend to participate in events created by other users in these groups. Combining the correlation between users and groups and the correlation between users in the group, the social interaction of user u about online group g is obtained. The preference score S _I (u, g) is shown in formula (14):

Among them, α∈[0,1] is used as the weight control parameter. In the social relationship network, it is equally important to set the preference association between the target user and the group and the association between the users in the group. The value of α here is set as the value of the experimental verification. is 0.5; Step 2-1-3, build the time factor preference model: The time factor of the event is an important preference factor that needs to be considered when calculating user preferences; the new event e that the user can choose to participate in is represented as a 7*24-dimensional event time vector

When a new event occurs in a specific time period of the week, the vector component value of the time period is set to 1, otherwise it is 0; in the time preference model, the user is represented as the user time according to the historical event records that the user participated in vector

As shown in formula (15):

Among them, E _u represents the set of historical events that the target user has participated in, and then calculate the cosine similarity s(u, e) between the user time vector and the new event time vector, as shown in Equation (16):

For a new event e, the user _ui ∈ U can obtain the similarity s( _ui , e) according to formula (16), and normalize the similarity to obtain the user’s time preference score for the event S _T ( _ui , e ), as shown in formula (17):

5. The personalized event recommendation method of merging theme matching and bidirectional preference as claimed in claim 4, wherein the calculating user preference score in step 2 specifically includes: For the geographic location preference model, the geographic location preference score is expressed by predicting the probability of the user participating in the event held at the location; for the social relationship preference model, the relationship between the target user and the group and the correlation with the users in the group are two factors. Calculate the social preference score of the target user; for the time factor preference model, a unified vector representation with two granularities of date and hour is constructed, and based on this, the similarity of the user-event pair is calculated as the time preference score of the target user; combined with These three single-factor preferences form a user preference perception model, and the linear combination of the three single-factor preferences is used to obtain the overall preference score S _user of user u to event e, as shown in formula (18):

Among them, _SG , _SI , and _ST represent the user's preference score on the three single factors of geographic location, social relationship, and time factor, respectively. 6. the personalized event recommendation method of fusion theme matching and bidirectional preference as claimed in claim 5, it is characterized in that, described in step 2 constructs event preference model respectively from event location popularity, event sponsor influence two Aspects to construct single-factor preferences for events, including: Step 2-2-1, build an event location popularity preference model: Calculate the popularity of geographic location according to user u and users in the online group g he joined to visit the place; First, define the popularity p( _le , u) of the event location _le with respect to the user u, as shown in equation (19):

Among them, the numerator m _l (u, _{le ) is the frequency of the user u participating in the activities held at the geographical location 1 e ,} and the denominator is the maximum frequency of the location that the user u has visited in the history _; The popularity p( _le , g) of group g is shown in formula (20):

Among them, the numerator represents the frequency of each user in group g participating in practical activities at location l, and the denominator is the maximum frequency of locations visited by group members in the past, from which the popularity of geographic location _le with respect to users in group g can be calculated. ; Combine p( _le , u) and p( _le , g) to define the total popularity of the location of the event to be recommended to the target user u as P( _le , u, g), as shown in formula (21) : P(le,u,g)= _αp ( _le ,u)+(1-α)p( _le ,g) (21) Step 2-2-2, construct the influence preference model of the event organizer: First, the influence degree of the event sponsor on the target user, the implicit preference of the event is expressed by the reputation or influence degree of the sponsor; the influence degree I(e, u) of the event on the user u is defined, as shown in Equation (22) ) as shown:

Among them, m _h (u, u _h ) represents the set of events organized by the organizer u _h that the user u has participated in, and E _h is the set of all events organized by the organizer u _h ; Second, the influence degree of the event organizer in the group, for the online group where the target user is located, the influence degree of the event in the group is represented by the proportion of the frequency of users participating in the group, and the influence degree of the user in the group is expressed as I( e, g) represent, as shown in formula (23):

Among them, U _g represents the set of users in the group u _h , m _h (u _i , u _h ) represents the set of events organized by the organizer u _h that the user _ui participates in, and E _h (g) represents that u _h is in the group u _h The set of events held in the event organizer; the comprehensive influence score I(e, u, g) of the event organizer is obtained by combining the influence of the event organizer on the target users and users in the group, as shown in formula (24): I(e,u,g)=αI(e,u)+(1-α)I(e,g) (24) 7. The personalized event recommendation method of fusion theme matching and bidirectional preference as claimed in claim 6, wherein the calculating event preference score in step 2 specifically includes: For new events that have not occurred, the preference of the event is expressed by calculating the event location popularity and event sponsor influence of the new event; for the constructed event location popularity P( _le , u, g) and event sponsor influence The force I(e, u, g) is linearly combined, and the preference score S _events of event e to user u is calculated, as shown in formula (25):

8. the personalized event recommendation method of fusion theme matching and bidirectional preference as claimed in claim 7, it is characterized in that, described in step 3 obtains user event bidirectional preference score, the theme matching degree score and bidirectional preference score are linearly weighted The final recommendation score of the user-event pair is obtained by combining, and the specific steps include: Step 3-1, find bidirectional preference for user-event pair: Assuming that the preference score weights of users and events are θ ₁ and θ ₂ respectively, the two-way preference score Su for user events is obtained by weighted fusion of the two _{, e} = θ ₁ S _user + θ ₂ S _events ; the problem of bidirectional preference score is converted into Find the weight vector of the two preference scores, and choose to use the implicit feedback as the training data to learn the weight vector; Select the learning algorithm BPR based on Bayesian maximum likelihood estimation to sort and learn the weights, and learn the correct sorting order of user-event pairs according to the user's implicit feedback data on events, so that the events that users participate in are ranked in new events or other events. Before; first, define the maximum posterior probability p(θ|R), as shown in Eq. (26): p(θ|R)∝p(R|θ)p(θ) (26) Among them, θ represents the weight vector, R represents the set of all user-event pairs, and p(R|θ) is defined as formula (27);

where R _u represents the user-event pair of user u, and p(ei > e _j ) represents the probability that event e _i ranks ahead of e _j for user _u , as shown in Equation (28): p(e _i >e _j |θ)=σ(s(u, e _i )-s(u, e _j )) (28) Among them, s(u, e) is the bidirectional preference score S _{u, e} ,

In order to make the optimization more convenient, assuming that θ obeys a normal distribution with a mean of 0, the final optimization objective function lnp(θ|R) can be derived from the expansion, as shown in Equation (29):

Among them, λ represents the regular term coefficient. The objective function is maximized through the implicit interactive feedback data of user events, and the optimal weight parameter vector is obtained. The stochastic gradient descent algorithm SGD is used to solve the optimization problem. The user-event pair of the target user is extracted to update the weight vector θ, and the update process is shown in formula (30):

Among them, α is the learning rate, s _ij =s(u, e _i )-s(u, e _j ); through the above learning process, the weight vector θ can be automatically obtained according to the user event preference score training set and hyperparameters α and λ , so as to obtain the bidirectional preference score S _{u, e} ; Step 3-2, combine topic matching and bidirectional preference to obtain the final recommendation score of the user-event pair: First, extract the event topic through the LDA topic model and obtain the topic matching score between the user and the event; secondly, according to the user event context information in the EBSN, the user and event preference models are respectively constructed, and the user event bidirectional preference score is obtained through the BPR learning algorithm. ; Finally, score the topic match

The final user-event pair recommendation score S _Rec is obtained by linearly weighted summation with user event bidirectional preference score _{Su, e} , as shown in formula (31):

Among them, γ is a weight parameter, which is usually set manually based on experience, and the optimal setting will be determined through experiments. 9. A realization system of the personalized event recommendation of fusion theme matching and bidirectional preference, it is used to realize the personalized event recommendation method of fusion theme matching and bidirectional preference as described in any one of claims 1-8, it is characterized in that That is, the implementation system includes: The document topic generation module is used to extract the topics of the user's historical events and new events, and calculate the topic distribution and word distribution of the events. One of the key factors is incorporated into the recommendation model for event recommendation; Build a user preference module, which is used to construct the user's single-factor preference from three aspects of geographic location, social relationship, and time factors, and weighted and fused the three single-factor preferences to obtain the user's overall preference; Build an event preference module, which uses the social influence of the event organizer in the group and the popularity of the geographical location where the event is held to represent the preference of the event; The user event bidirectional preference scoring module uses the sorting learning algorithm to solve the weight parameters of the user preference score and the event preference score, and obtains the user event bidirectional preference score; The final recommendation scoring module for user-event pairs is used to linearly weight the topic matching score and the bidirectional preference score to obtain the final recommendation score for the user-event pair. 10. The implementation system for combining theme matching and bidirectional preference personalized event recommendation as claimed in claim 9, wherein the user preference module comprises a geographic location preference module, a social relationship preference module and a time factor preference module, wherein The event preference module includes the event location popularity preference module and the event sponsor influence preference module, wherein: The geographic location preference module is used to represent the geographic location preference score by predicting the probability that the user participates in an event activity held in a certain geographic location; The social relationship preference module is used to calculate the social preference score of the target user from two aspects: the relationship between the target user and the group and the correlation with the users in the group; The time factor preference module is used to construct a unified vector representation of two granularities of date and hour, and calculate the similarity of the user-event pair as the time preference score of the target user; The event location popularity preference module is used for when a new event is recommended, the location of the event is an important selection basis for interested users, which is called the popularity of geographic location in the user group, considering the influence of event geographic location. Popularity can more accurately calculate the attractiveness of events to users; The event sponsor influence preference module is used to improve the accuracy of recommendation according to the influence of the event sponsor in the group where the target user is located, from the influence of the event sponsor on the target user and the event sponsor in the group. The influence is calculated from two aspects.

Images