CN111428127B

CN111428127B - Personalized event recommendation method and system integrating theme matching and bidirectional preference

Info

Publication number: CN111428127B
Application number: CN202010069262.9A
Authority: CN
Inventors: 钱忠胜; 杨家秀; 朱懿敏
Original assignee: Jiangxi University of Finance and Economics
Current assignee: Jiangxi University of Finance and Economics
Priority date: 2020-01-21
Filing date: 2020-01-21
Publication date: 2023-08-11
Anticipated expiration: 2040-01-21
Also published as: CN111428127A

Abstract

The invention discloses a personalized event recommendation method and system integrating theme matching and bidirectional preference. Firstly, extracting topic information of an event and a historical event participated by a user by using a document topic generation model LDA, and calculating topic matching degree of the user and the event; secondly, considering the social network recommendation based on the event from the two-way angles of the user and the event, constructing preference models of the user and the event, respectively obtaining user preference scores and event preference scores, and more completely mining preference relations from the two angles of the user and the event; finally, the user-event pair matching degree is combined with the user event bi-directional preference linear weighted combination to obtain the final user-event pair comprehensive score, and the ordered TOP-K user-event pairs are used as recommendation results. The performance of the recommendation algorithm of the scheme is superior to that of the traditional recommendation scheme, and the personalized preference of the user can be well predicted, so that the purpose of personalized recommendation is achieved.

Description

Personalized event recommendation method and system integrating theme 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 theme matching and bidirectional preference.

Background

With the rapid development of internet and computer technology, in recent years, traditional social networks have also developed towards different innovations, and some special types of novel social networks, such as Location-based social networks (Location-Based Social Network, LBSN), are formed, social networks mainly forming social relationships according to geographic sign-in information of users, and other complex heterogeneous social networks combined on-line and off-line, namely event-based social networks (Events-Based Social Network, EBSN), are distinguished from friend relationships established between acquaintances in traditional social networks, in which users establish interpersonal relationships through social activities, and users join in interest groups on-line and collective social activities off-line according to own interests or common points.

In the rapid development process of the event-based social network, more and more users choose to participate in social activities in the event-based social network, on the event-based social network platform, the users can join various online groups, and an organizer or users in the group can initiate and participate in any offline social activity, such as a party, hiking, sports activities, singing, and the like, and share information with other users.

The event-based social network may provide the user with online-to-offline integrated social services that help the user initiate and formulate a personalized event participation plan. Users form online group relationships by common interests on the line and initiate offline meeting events on the line, event-based social networks possess broader social attributes than location-based social networks, and existing work has shown that event social networks possess better recommendation characteristics than traditional social networks in recommendation systems.

Most current event-based social network recommendations are based primarily on user unidirectional angle extraction feature preferences, and while considering the social impact of event sponsors, the potential attractiveness to events is not adequate. On the other hand, the influence of the theme factors only takes the event theme as one of the recommended factors, and the user theme factors and the matching degree of the user theme factors and the event theme are less considered.

Disclosure of Invention

In view of the foregoing, it is desirable to provide a personalized event recommendation method and system that combines several types of primary context information to calculate user preferences and event potential preferences, and ultimately fuses the topic matching and bi-directional preferences of the user-event with the topic matching degree.

A personalized event recommendation method integrating theme matching and bidirectional preference comprises the following steps:

firstly, extracting topic information of an event by using a document topic generation model LDA, obtaining user topic information according to a historical event record participated by a user, calculating topics of a new event and a user historical event, and calculating a topic matching degree score of a user-event pair by adopting a JS divergence algorithm;

respectively constructing a user preference model and an event preference model, and respectively calculating a user preference score and an event preference score;

thirdly, learning weight parameters of the user preference scores and the event preference scores by using a Bayesian personalized ranking algorithm BPR to obtain user event bidirectional preference scores, linearly weighting and combining the topic matching degree scores and the bidirectional preference scores to obtain final recommendation scores of the user-event pairs, and recommending the top K events after ranking to the user.

Further, the document topic generation model LDA in the first step has a three-layer generation type bayesian network structure comprising a document, a topic and a word, wherein the document-topic and the topic-word are subjected to polynomial distribution; each document selects a topic with a certain probability, and a word is selected from the topic with a certain probability, and topics in any document conform to the Dirichlet distribution, and relationships among texts are discovered through the distribution.

Further, in the first step, the calculating the theme of the new event and the user history event, and calculating the theme matching degree score of the user-event pair by adopting a JS divergence algorithm, which specifically includes the following steps:

step 1-1, all event description contents are formed into a document setDAnd removing stop words, and collecting the documentsDInputting a document theme generation model LDA, and respectively obtaining theme distribution of each event;

removing stop words and punctuation marks from all event contents, and regarding the document contents with noise interference words removed as a set of all documentsDInput into LDA topic model to generate documentAssociative distribution of topics and words->As shown in formula (1):

(1)；

two unknown parameters in the model are then estimated using the Gibbs sampling method: event topic distributionAnd subject word distribution->；

Step 1-2, calculating the topic distribution similarity between the historical event and the new event of the target user according to a JS divergence algorithm;

the topic distribution of all events has been generated according to equation (1)Given event-> andRespectively have theme distributionThe JS divergence +.>As shown in formula (2):

(2)；

wherein ,，represent KL divergence to describe two probability distributions +. > andThe difference between the two is calculated as shown in formula (3):

(3)；

the event obtainable by combining formula (2) and formula (3) andSubject similarity of->As shown in formula (4):

(4)；

wherein the topic similarity of the eventsThe value of (1) is at [0,1 ]]In (2), the closer the value is to 1, the higher the event similarity is;

step 1-3, averaging the similarity of all the historical events of the target user to obtain a subject matching degree score of the user and the new event;

to be used forRepresenting the historical event number of the target user, taking the average value of all the similarity of the target user +.>As a topic matching degree score for the user and the new event, as shown in formula (5):

(5)；

according to the constructed topic matching model, finallyTo measure the topic matching relationship between the target user and the new event.

Further, the building user preference model in the second step builds single-factor preferences of the user from three aspects of geographic location, social relationship and time factor, and specifically includes:

step 2-1-1, constructing a geographic location preference model:

the geographic position preference model calculates the probability that a target user will participate in holding an event at the position, a kernel density estimation KDE method is adopted to model the two-dimensional geographic position distribution of the event in which the user participates, and the event participation probability after normalization is used to represent the preference degree of the user to the geographic position. Longitude and latitude coordinates of event geographic position Lx, Ly) Representing a set of places where a user historically participated in an eventL(u) Representation, then about the useruKDE function of (F)As shown in formula (6):

(6)；

wherein ,l _i =(Lx _i ,Ly _i ) ^T a two-dimensional vector representing the longitude and latitude coordinates of the event location,m _l (u,l _i ) Representing a useruParticipating geographic locationsl _i The frequency of the holding-up of the event is at the process,σrepresenting the size of the neighborhood window (bandwidth),Nthe number in the position sample is indicated,K(•) The Gaussian kernel function is represented, and the definition form is shown as a formula (7):

(7)；

the combination of formula (6) and formula (7) may define a useruThe participation will be at the positionlProbability of event held is calculated as shown in formula (8):

(8)；

normalizing the probability to obtain preference scores of users about geographic positionsAs shown in formula (9):

(9)；

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

step 2-1-2, constructing a social relation preference model:

in a user social relationship network, a user can join at least one or more interest groups on line, select event activities which participate in different groups to issue, and judge social relationship preference of the user through on-line same-group relationship of the user, wherein the same-group relationship mainly comprises two interaction relationships;

first, the relevance of users to groups is defined as the interactive relationship between users and all groups they belong to and between users and events created within groups, to G(u) Representing a useruSet of groups to which the participating events belong, then the relevance of the user to the groupCan be represented by the formula (10):

(10)；

wherein ,m _p (u,g) Representing users in a group of usersuEvent activity sets that were attended;

second, intra-group user relevance, which is defined by the similarity of friends in the group where the target user is located, calculates the similarity of the target user and the intra-group userAs shown in formula (11):

(11)；

wherein ,sim(u _i ,u _j ) Representing users in the same groupu _i And a useru _j The similarity is shown as a formula (12);

(12)；

will bes(u,g) Normalized toAs shown in formula (13):

(13)；

combining the two interaction relations, users belonging to the same group tend to participate in events created by other users in the groups, and the correlation between the users and the group and the correlation between the users in the group are integrated to obtain the usersuWith respect to on-line teamsgSocial preference scoring of (2)As shown in formula (14):

(14)；

wherein ,as a weight control parameter, in a social relation network, setting the preference association of a target user and a group to be equally important with the association between users in the group, and verifying the preference association by experiments>The value of (2) is set to 0.5;

step 2-1-3, constructing a time factor preference model:

the time factor of the event is an important preference factor to be considered when calculating user preferences; new event with optional participation of user eRepresented as a 7 x 24 dimensional event time vectorWhen a new event occurs in a certain specific time period of the week, setting the vector component value of the time period to be 1, otherwise, setting the vector component value to be 0; representing the user as a user time vector +.>As shown in formula (15):

(15)；

wherein ,E _u representing a set of historical events in which the target user participated, and then calculating cosine similarity between the user time vector and the new event time vectorAs shown in formula (16):

(16)；

for new eventsUser->The similarity can be determined according to formula (16)>Normalizing the similarity results in a time preference score for the user for the event +.>As shown in formula (17):

(17)；

further, the calculating the user preference score in the second step specifically includes:

for the geographic position preference model, representing a geographic position preference score by predicting the probability of a user participating in event activities held at the position; calculating social preference scores of the target users from two aspects of the relationship between the target users and the group and the relevance between the target users and the users in the group for the social relationship preference model; for the time factor preference model, a unified vector representation of two granularities of date and hour is constructed, and the similarity of the user-event pairs is calculated as the time preference score of the target user based on the unified vector representation; combining the three single-factor preferences to form a user preference perception model, and obtaining the user by linearly combining the three single-factor preferences uFor eventseOverall preference scoring of (a)As shown in formula (18):

(18)；

wherein ,the preference scores of the users on three single factors of geographic position, social relationship and time factor are respectively represented.

Further, the building event preference model in the second step builds single factor preferences of the event from two aspects of event location popularity and event sponsor influence, which specifically includes:

step 2-2-1, constructing an event location popularity preference model:

according to the useruAnd the on-line team to which it is addedgCalculating popularity of the geographic location based on the frequency of access to the location by the user;

first define event geographyPosition ofl _e In relation to usersuPopularity of (2)As shown in formula (19):

(19)；

wherein the moleculeFor usersuParticipating geographic locationsl _e Frequency of holding activities is treated as useruThe maximum frequency of historic visited locations; likewise, a geographic location is definedl _e In relation to usersuGroup ofgPopularity of->As shown in formula (20):

(20)；

wherein the molecules represent a panelgEach user in the locationlThe frequency of participation in the practice activities is the maximum frequency of the historic visited positions of the members of the group, thereby calculating the geographic positionl _e With respect to the groupgPopularity of users in (a); bonding of andDefining a holding location for an event to be recommended to a target useruThe total popularity of->As shown in formula (21):

(21)；

step 2-2-2, constructing an event sponsor influence preference model:

first, the influence of the event sponsor on the target user selects implicit preference for representing the event by the reputation or influence of the sponsor; defining events to usersuInfluence degree of (2)As shown in formula (22):

(22)；

wherein ,representing a useruParticipating sponsorsu _h The set of events held is then displayed,E _h is a sponsoru _h All event sets held;

second, the influence of event sponsors in the group is expressed by the ratio of the frequency of user participation for the on-line group of target users, and the influence of the users in the group is expressed by the ratio of the frequency of user participationThe expression is as shown in formula (23):

(23)；

wherein ,U _g representation teamUser set in->Representing user +.>Participating sponsors->Event collection held,/->Representation->In the group->A set of events held in the middle; obtaining comprehensive influence degree scores of event sponsors by combining influence degree of the event sponsors on target users and on users in the group>As shown in formula (24):

(24)；

further, the calculating the event preference score in the second step specifically includes:

For new events which do not occur, representing the preference of the event by calculating the event location popularity and event sponsor influence of the new event; popularity for constructed event locationsAnd event sponsor influence +.>Linear combination, calculation of eventseFor the useruPreference score +.>As shown in formula (25):

(25)；

further, the obtaining a bidirectional preference score of the user event in the third step, and linearly weighted combining the topic matching degree score and the bidirectional preference score to obtain a final recommendation score of the user-event pair, which specifically includes:

step 3-1, two-way preference is made for user-event pairs:

assume that the preference scoring weights for the user and event are respectively andThe user event bidirectional preference score +.>The method comprises the steps of carrying out a first treatment on the surface of the Converting the bi-directional preference scoring problem into weight vectors for the two preference scores, and selecting implicit feedback to be used as training data for learning the weight vectors;

selecting a learning algorithm BPR based on Bayesian maximum likelihood estimation to perform sequencing learning on weights, and learning a correct sequencing order of user-event pairs according to implicit feedback data of the user on the events so that the events participated by the user are ranked before new events or other events; first, define the maximum posterior probability As shown in formula (26):

(26)；

wherein ,θthe weight vector is represented by a weight vector,Rrepresenting a set of all user-event pairs,the definition is shown as a formula (27);

(27)；

wherein, in the formulaRepresenting user +.>User-event pairs of->Representing +.>Event->Arranged at->The former probability is shown in formula (28):

(28)；

wherein ,i.e. bi-directional preference score->，The method comprises the steps of carrying out a first treatment on the surface of the For more convenient optimization, assume +.>Obeying the normal distribution with the mean value of 0, developing and deducing to obtain a final optimized objective function +.>As shown in formula (29):

(29)；

wherein ,representing regular term coefficients, maximizing optimization objective function through implicit interactive feedback data of user eventsObtaining an optimal weight parameter vector; solving the optimization problem by adopting a random gradient descent algorithm SGD, randomly extracting user-event pairs of target users from a training set in an iterative process to update weight vectors +.>The update process is as shown in equation (30):

(30)；

wherein ,is learning rate (I/O)>The method comprises the steps of carrying out a first treatment on the surface of the Through the learning process, the training set and the super parameter ++can be automatically scored according to the user event preference> andObtaining a weight vector->Thereby obtaining a bi-directional preference score->；

Step 3-2, combining topic matching and bidirectional preference to obtain final recommendation scores of the user-event pairs:

Firstly, extracting an event topic through an LDA topic model and obtaining a topic matching degree score of a user and the event; secondly, respectively constructing preference models of users and events according to user event context information in the EBSN, and obtaining bidirectional preference scores of the user events through a BPR learning algorithm; finally, scoring the topic matching degreeBi-directional preference score with user event->Linear weighted summation to get final user-event pair recommendation score +.>As shown in formula (31):

(31)；

wherein ,for weight parameters, typically manually set empirically, the optimal setting will be determined experimentally.

And a realization system for fusing topic matching and bidirectional preference personalized event recommendation, which is used for realizing the method for fusing topic matching and bidirectional preference personalized event recommendation according to any one of the above, the realization system comprises:

the document theme generation module is used for extracting the theme of the user history event and the new event, calculating the theme distribution and the word distribution of the event, expressing the theme matching degree by using the theme similarity between the user history event and the new event, and fusing the theme matching degree into a recommendation model as one of the recommended key factors so as to recommend the event;

The user preference module is used for constructing single-factor preferences of the user from three aspects of geographic positions, social relations and time factors, and weighting and fusing the three single-factor preferences to obtain overall preferences of the user;

constructing an event preference module, and representing the preference of the event by using social influence of event sponsors in the group and popularity of geographic positions of event sponsors in the group;

the user event bidirectional preference scoring module is used for solving the weight parameters of the user preference scores and the event preference scores by using a sequencing learning algorithm to obtain user event bidirectional preference scores;

and the final recommendation scoring module is used for linearly weighting and combining the topic matching degree scores and the bidirectional preference scores to obtain final recommendation degree scores of the user-event pairs.

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

the geographic position preference module is used for representing geographic position preference scores by predicting the probability of a user participating in event activities held at a certain geographic position;

The social relation preference module is used for calculating social relation preference scores of the target users from two aspects of relation between the target users and the group and correlation between the target users and users in the group;

the time factor preference module is used for constructing unified vector representation of two granularities of date and hour and calculating the similarity of the user-event pairs as the time preference score of the target user;

the event position popularity preference module is used for selecting important basis of holding places for interested users when recommending new events, namely popularity of geographic positions in a user group, and the attraction of the events to the users can be calculated more accurately by considering the popularity of the geographic positions of the events;

the event sponsor influence preference module is used for improving the recommendation accuracy according to the influence of the event sponsor on the group where the target user is located, and calculating the influence of the event sponsor on the influence of the target user and the influence of the event sponsor on the group.

In the personalized event recommendation method and system based on fusion of the topic matching and the bidirectional preference, firstly, the topic information of the event is extracted by utilizing the document topic generation model LDA, the user topic information is obtained according to the historical event record of the user participation, the topic matching degree of the user and the event is calculated to be used as an important recommendation factor in the recommendation model, and the topic factor can better represent the feature preference; secondly, considering the social network recommendation based on the event from the two-way angles of the user and the event, constructing preference models of the user and the event, respectively obtaining user preference scores and event preference scores, and more completely mining preference relations from the two angles of the user and the event; finally, the user-event pair matching degree is combined with the user event bi-directional preference linear weighted combination to obtain the final user-event pair comprehensive score, and the TOP K (namely TOP-K) user-event pairs after sequencing are used as recommendation results. A large number of experiments are carried out on the Meetup real data set, and the comparison is carried out with other event recommendation algorithms, so that the performance of the software recommendation algorithm is superior to that of the traditional recommendation scheme, the personalized preference of a user can be well predicted, and the purpose of personalized recommendation is achieved.

Drawings

FIG. 1 is a diagram of an overall recommendation fusion framework of a personalized event recommendation method and system fusing topic matching and two-way preference in accordance with an embodiment of the present invention.

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

Detailed Description

In this embodiment, a personalized event recommendation method integrating 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.

Referring to fig. 1 and fig. 2, a personalized event recommendation method and system for fusing topic matching and bidirectional preference are shown.

The technical details related to the personalized event recommendation system integrating theme matching and bidirectional preference of the software are specifically explained. The main idea is that firstly, the topics of new events and user history events are calculated through an LDA topic model, the topic matching degree of user-event pairs is calculated by adopting cosine similarity, and a user preference model and an event preference model are respectively constructed. Wherein, the user preference model calculates the comprehensive preference score of the user from three aspects of time, geography and social relationship. The event preference model represents event potential preference scores based on the popularity of new events in the geographic locations of the target user group and the degree of social impact within the group of sponsors. The user event bi-directional preference scores are then obtained by learning the weighting parameters of the user preference scores and the event preference scores using a bayesian personalized ranking algorithm (Bayesian Personalized Ranking, BPR). Finally, the final recommendation degree scores of the user-event pairs are obtained through linear weighted fusion with the topic matching degree, and the ordered TOP-K events are recommended to the user. The software matches the user with the event topic, calculates the user preference and the event potential preference by combining several types of main context information, and finally fuses the topic matching degree and the user-event bidirectional preference to conduct event recommendation.

1. Recommendation framework integrating LDA topic matching and user event bidirectional preference

Based on the current existing work, an event recommendation scheme combining user-event pair topic matching and user-event pair bidirectional preference is proposed based on geographic position information, time information, social relations and other relevant user event context information in the EBSN. In the scheme, the influence of the theme matching degree, the user preference and the event preference on the event recommendation is considered respectively, and the factors are fused to effectively recommend the interest event to the user. The overall framework of the recommendation model is shown in fig. 1, and the specific recommendation process is as follows:

1) According to the description document of the event in the EBSN, the LDA theme model is utilized to calculate the history event theme of the new event and the target user, the theme of the user history event is used for representing the theme of the user, then the semantic similarity of the event and the user theme distribution is calculated, and the matching degree score of the user-event theme is obtained.

2) The user preference score and the event preference score are calculated, the preference scores are calculated from three aspects of geographic position, social relation and time for the user preference and are linearly fused, the event preference is represented by popularity of the event holding geographic position and social influence of the event sponsor, and the event preference score is obtained by linearly fusing the event preference. It should be noted that when calculating the popularity of the geographic location and the influence of the sponsor on the event, only the group and the users in the group where the target user is located are ignored for the other users and the association of the group, so as to improve the recommendation performance and reduce the calculation complexity.

3) The matching degree score of the user-event theme and the preference score of the user on the event and the preference score of the event on the user are obtained through the calculation. Firstly, learning weights of user preference scores and event preference scores by using a Bayesian personalized ranking algorithm, obtaining bidirectional preference scores by fusing the preference scores of the users and the events according to the weights, finally obtaining final user-event pair recommendation scores by linearly combining the topic matching degree scores and bidirectional preference score information, and recommending TOP-K events with highest scores to the users.

2. Topic matching model based on LDA

There is a clear topic semantic similarity relationship between users and events in an event social network, and users typically choose to participate in a certain class of events of interest, which typically have similar attributes and topics. The theme of the application event in the recommendation can better capture the user and the preference of the event, the theme of the historical event attended by the user is used for representing the user theme, the new event theme distribution and the word distribution are calculated, the theme similarity between the historical event of the user and the new event is used for representing the theme matching degree, and the theme matching degree is used as one of the key factors of the recommendation and is fused into a recommendation model for event recommendation.

When two documents have the same features such as topic, it is difficult to distinguish the two objects using TF-IDF (Term Frequency-Inverse Document Frequency) algorithm, so a bayesian-based LDA topic model is selected to calculate document topic distribution and word distribution. The LDA topic model is a Bayesian probability model for computing a document topic distribution for clustering potential topics for a document and generating document topics. The key idea is that each document selects a certain topic with a certain probability, and selects a certain word from the topics with a certain probability, and the topics in any document are considered to accord with the Dirichlet distribution, and the relation among the texts can be discovered through the distribution. LDA consists of a three-layer generative bayesian network structure containing documents, topics, and words, both document-topic and topic-word obeying polynomial distributions. The LDA topic model generation process is shown in fig. 2.

Given set of documentsIn FIG. 2-> andRespectively represent document->Is a priori Dirichlet distribution of topic distribution and word distribution,/->The superparameter of the subject prior distribution and the word prior distribution given empirically,kis the number of topics of the previously specified document set,N _m representation document- >Is used to determine the number of words in the word,Mis the number of documents in the document set. For documents->LDA is based on a priori knowledge +.>Determining topic distribution of a document>Then distribute->Extracting a themezYet according to a priori knowledge->Determining word distribution of a current topic->Then from the main partQuestions (questions)zCorresponding word distribution->Extracting a word-> ，Repeating the above processN _m The document +.>. In this process, the Gibbs sampling method can be used to solve the document +.>Is a topic distribution of (1).

According to the topic similarity between the LDA computing user and the event, converting the text content into semantic features, and computing topic distribution for each event by using an LDA topic model. The event content mainly comprises a title and a description document, and also comprises information such as time, holding place and the like, and the event theme can be extracted through the event content. In contrast, the user can choose to set the interest tag to represent the preference, however, many users cannot set the interest tag or the self-introduction and other contents, the user content lacks document information and faces the problem of extremely sparse data, and at the moment, no available features represent the user theme, so that the theme of the historical event in which the user participates is chosen to express the user theme more accurately, and the problems of sparse data and blank tags are avoided. Removing stop words and punctuation marks from all event contents, and regarding the document contents with noise interference words removed as a set of all documents DInput into LDA topic model, documents are generated according to the generation process described aboveJoint distribution of topics and wordsAs shown in formula (1). Then using Gibbs sampling method to estimate two unknown parameters in the model, namely event topic distribution +.>And subject word distribution->。

(1)

After obtaining the topic distribution and the word distribution of the event document through the LDA process, the JS divergence (Jensen Shannon divergence) method is utilized to calculate the similarity between the events according to the topic distribution of the events. The JS divergence is a variation based on KL divergence (Kullback-Leibler divergence), is symmetrical, solves the problem of asymmetry of KL divergence, and can better measure the similarity of two probability distributions. The topic distribution of all events has been generated according to equation (1)Given an event andRespectively have theme distribution->The JS divergence +.>As shown in formula (2).

(2)

wherein ,，represent KL divergence to describe two probability distributions +.> andThe difference between the two is shown in a formula (3).

(3)

The event obtainable by combining formula (2) and formula (3) andSubject similarity of->As shown in formula (4).

(4)

Topic similarity of eventsThe value of (1) is at [0,1 ]]In (2), the closer the value is to 1, the higher the event similarity is. It has been mentioned above that the topic similarity between the new event and the user history event is taken as the topic similarity between the user and the event, and the user often participates in the event a plurality of times, and a plurality of topic similarities exist between the new event, so as +. >Representing the historical event number of the target user, taking the average value of all the similarity of the target user +.>As a topic matching degree score for the user and the new event, as shown in formula (5).

(5)

Algorithm 1 describes the topic matching process of computing user-event pairs by LDA topic model, whereWord distribution representing topic->Representing the distribution of the subject matter of the document,Dir() The Dirichlet distribution is represented by the expression,Mult() The distribution of the polynomials is represented,Poiss() Representing poisson distribution. />

Algorithm 1 gives a process of solving the user-event matching degree score of the topic by using the LDA topic model and the JS divergence algorithm. Firstly, forming all event description contents into a document set, removing stop words, and respectively obtaining the topic distribution (line 2 to line 11) of each event as the input of an LDA model; calculating the topic distribution similarity (12 th line to 14 th line) between the historical event and the new event of the target user according to the JS divergence algorithm; and finally, averaging the similarity of all the historical events of the target user to obtain the subject matching degree scores (15 th line to 16 th line) of the user and the new event.

3. User-based preference model

Feature learning is typically performed for user preferences from relevant context information of the user, and the learned feature information is represented as user preferences. The single-factor preference of the user is built from three aspects of geographic factors, social relations and time factors, and the three single-factor preferences are weighted and fused to obtain the overall preference of the user.

3.1 Geographic location preference

The geographic position preference model calculates the probability that a target user will participate in holding an event at the position, a KDE (Kernel Density Estimation ) method is adopted to model the two-dimensional geographic position distribution of the event participated by the user, and the event participations after normalization are probably usedThe rate indicates the user's preference for the geographic location. Longitude and latitude coordinates of event geographic positionLx, Ly) Representing a set of places where a user historically participated in an eventL(u)Representation, then about the useruKDE function of (F)As shown in formula (6).

(6)/>

wherein ,l _i =(Lx _i ,Ly _i ) ^T a two-dimensional vector representing the longitude and latitude coordinates of the event location,m _l (u,l _i ) Representing a useruParticipating geographic locationsl _i The frequency of the holding-up of the event is at the process,σrepresenting the size of the neighborhood window (bandwidth),Nthe number in the position sample is indicated,K(•) The gaussian kernel function (Gaussian kernel function) is represented by the following formula (7).

(7)

The combination of formula (6) and formula (7) may define a useruThe participation will be at the positionlThe probability of event activity held is shown in formula (8).

(8)

Normalizing the probability to obtain preference scores of users about geographic positionsAs shown in formula (9).

(9)

3.2 Social relationship preferences

In a user social relationship network, users typically join at least one or more interest groups online and may choose to participate in event activities published by different teams. In these group relationships, the user usually selects a preference group with the most interesting self to participate in, and members in the same group generally have the same interest, so that social relationship preferences of the user can be considered through the online same group relationship of the user, and mainly comprise two interaction relationships.

1) Correlation of users with groups. I.e. the interactive relationship between the user and all the groups to which they belong and between the user and the events created within the groups. To be used forG(u) Representing a useruSet of groups to which the participating events belong, then the relevance of the user to the groupCan be represented by formula (10).

(10)

wherein ,m _p (u,g) Representing users in a group of usersuEvent activity sets that were attended.

2) Intra-group user relevance. The intra-group user relevance is defined by the friend similarity in the group of the target user, and the similarity between the target user and the intra-group user is calculatedAs shown in formula (11).

(11)

wherein ,sim(u _i ,u _j ) Representing users in the same groupu _i And a useru _j The similarity is shown in the formula (12).

(12)

Finally, wills(u,g) Normalized toAs shown in formula (13).

(13)

Combining these two interactions, users belonging to the same or similar groups tend to participate in events created within those groups, and combining the user's relevance to the group and the intra-group user relevance to arrive at a useruWith respect to on-line teamsgSocial preference scoring of (2)As shown in formula (14).

(14)

wherein ,as a weight control parameter, in a social relation network, the preference association of a target user and a group is considered to be equally important as the association between users in the group, and experimental verification proves that +. >The value of (2) is set to 0.5.

3.3 Time preference

The time factor of an event is another important preference factor that needs to be considered when calculating user preferences. For different users to have different preferences in selecting to engage in event activities, some users may prefer to engage in activities at night, while others may prefer to engage in activities in the morning, or prefer different points in time on weekdays or weekends. In reality, the time is periodic, mainly with 7 days per week and 24 hours per day, and for the user to choose to engage in activities on a day of the week and for hours of the day, two different levels of user time preferences are formed. We represent the user's time preference by combining the user selections at two levels of granularity.

If the user chooses to engage in an activity during a certain time period of a day of the week, which may indicate an implicit time preference of the user, the user may choose to engage in an event activity again during the same time period next. To intuitively represent this implicit preference in a unified way, we will choose to have the user opt-in to a new eventeRepresented as a 7 x 24 dimensional event time vector. When a new event occurs in a certain period of time of the week, the vector component value for that period of time is set to 1, otherwise to 0. Therefore, the user can be represented in the time preference model as a user time vector according to the history of user participation >As shown in formula (15).

(15)

wherein ,E _u representing a set of historical events in which the target user participated, and then calculating cosine similarity between the user time vector and the new event time vectorAs shown in formula (16).

(16)

For new eventsUser->The similarity can be determined according to formula (16)>NormalizationObtaining a time preference score of the user for the event by transforming the similarity>As shown in formula (17).

(17)

3.4 User fusion preference scoring

The preference scores of the user with respect to geographic location, social relationship, and time are calculated separately from the previous modeling of the user's one-factor preference model from three aspects. For a geographic location, representing a geographic location preference score by predicting a probability of a user engaging in an event activity held at the location; for the social relationship, calculating social preference scores of the target users from the relationship between the target users and the group and the relevance between the target users and the users in the group; for time preference, then, a unified vector representation of both date and hour granularity is constructed, and based thereon, the similarity of the user-event pairs is calculated as the time preference score for the target user. Combining the three single-factor preferences to form a user preference perception model, and obtaining the user by linearly combining the three single-factor preferences uFor eventseOverall preference scoring of (a)As shown in formula (18).

(18)

wherein ,the distribution represents the preference scores of the user over three factors, geographic location, social relationship, and time. Algorithm 2 describes the process of calculating the user preference score. />

Algorithm 2 gives a process of solving the user comprehensive preference score by combining the preferences of the user in three factors of geographic position, social relationship and time. Predicting the probability that the user possibly participates in an event held at a specific position through a kernel density estimation algorithm, normalizing the probability and representing the geographic preference of the user (line 3); calculating social associations of users with online teams and members within a team according to formulas (10) and (13) to represent social preferences (line 5 to line 11); representing the new event and the user history event as time vectors, and calculating cosine similarity of the new event and the user history event to represent time preference of the user (line 4); finally, the three preference values are linearly combined to obtain the total preference score of the user (line 13 to line 14).

4. Event-based preference model

For event preferences, consider learning from event sponsors and event ontology information. Because the event lacks active personalized context information compared to the user, it does not have information such as history, personalized tags, etc. for a new event, the preference of the event is expressed in terms of social impact of the event sponsor in the group, and popularity of the geographic location of the event sponsor in the group.

4.1 Event location popularity

The geographic location at which an event is held is one consideration in the user's choice of whether to attend an event. For a group of users on a line, which are generally of the same interest, there may be multiple users opting to participate in the same event activity, so for a new event recommendation, the place of holding may be an important basis for the interested users, and this relationship is called popularity of geographic location in the group of users. Considering popularity of geographic locations of events in a model that calculates event preferences enables more accurate calculation of attractions of events to users. According to the useruAnd the on-line team to which it is addedgThe popularity of geographic locations is calculated by the frequency of site access by users in (a) to (b).

First define the geographic location of an eventl _e In relation to usersuPopularity of (2)As shown in formula (19).

(19)/>

Wherein the moleculeFor usersuParticipating geographic locationsl _e Frequency of holding activities is treated as useruThe maximum frequency of historic visited locations. Likewise, geographic locations may be definedl _e In relation to usersuGroup ofgPopularity of (2)As shown in formula (20).

(20)

Wherein the molecules represent a panelgEach user in the location lThe frequency of participation in the practice activities is the maximum frequency of the historic visited positions of the members of the group, thereby calculating the geographic positionl _e With respect to the groupgPopularity of users in (a). Bonding of andMay define the location of the event to be recommended to the target useruThe total popularity of->As shown in formula (21).

(21)

4.2 Event sponsor influence

In the event social network, the initiator of each event activity is also an ordinary user on the network, and a general sponsor initiates an activity to obtain better reaction, so that the user participated before is highly likely to opt to participate in the activity held again when other new activities are initiated next time. While the event to be recommended is a completely new event that has not yet occurred for each user, the sponsor of the event may be an active sponsor of this type of event, and may have been sponsored multiple events before, which provides more auxiliary recommendation information for solving the cold start problem that exists in event recommendation. As can be seen, the impact of event sponsors in the user population within the team is an important feature of event preferences, and the software improves the accuracy of the recommendation based on the impact of event sponsors in the target user population. Its influence can be considered from the following two aspects.

1) The impact of the event sponsor on the target user. The scoring information of the event by the user does not exist in the event social network, the influence of the sponsor and the event cannot be intuitively represented, and the scoring of the event at the end of the life cycle of the event has no practical meaning, and the implicit preference of the event is represented by the credibility or influence of the sponsor because the new event held later cannot be influenced. First define event to useruInfluence degree of (2)As shown in formula (22).

(22)

wherein ,representing a useruParticipating sponsorsu _h The set of events held is then displayed,E _h is a sponsoru _h All events held are collected.

2) Influence of event sponsors in the panel. For a group of online users, the impact of an event in the group can be similarly represented by the ratio of frequencies in which the users participateInfluence of users in groupThe expression is shown in the formula (23).

(23)

wherein ,U _g representation teamUser set in->Representing user +.>Participating sponsors->Event collection held,/->Representation->In the group->A set of events held in the computer. The comprehensive influence degree score of the event sponsor can be obtained by combining the influence degree of the event sponsor on the target user and the users in the group >As shown in formula (24).

(24)

4.3 Event potential preference scoring

For new events that do not occur, the present software sets the two key factors that attract users to attend as the geographic location and the influence of the sponsor. By passing throughThe geographic location popularity of new events and the social impact of their sponsors are calculated to represent the event's preferences. In order to reduce the computational complexity and avoid the interference and influence of weak related data, the popularity of the geographic location and the social influence of the sponsor on the event are limited to the group where the target user is. The rest of the users or teams are here assumed to have zero relevance, with no impact on event preferences. For the event geographic position popularity constructed by the methodAnd sponsor influenceLinear combination to determine eventseFor the useruPreference score +.>As shown in formula (25).

(25)

Algorithm 3 details the process of calculating event potential preference scores from event location popularity and sponsor impact.

Algorithm 3 presents a process for solving for event potential preference scores based on event geographic location popularity and sponsor influence. For the group of target users, calculating popularity of the event geographic position to the users and the group according to the formula (19) and the formula (20), and combining the popularity and the popularity to represent the total popularity of the event geographic position (3 rd line to 8 th line); similarly, the influence of the event sponsor on the users and the group is obtained by the formula (22) and the formula (23) (line 9 to line 13), and the influence of the event sponsor is expressed by combining the two; finally, the potential preference scores for the event are obtained by a linear combination of location popularity and sponsor impact (line 17).

5. Recommendation algorithm integrating topic matching and user event bidirectional preference

The topic distribution of the user and the event is solved by utilizing the LDA topic model, and the topic matching degree of the user-event pair is calculated according to the topic distribution; and then, a characteristic preference scoring model is constructed for the user and the event, and the user preference score and the event preference score are respectively obtained. The method comprises the steps that topic matching and user event preference are fused to obtain final recommendation scores, and firstly, a ranking learning algorithm is utilized to solve weight parameters of the user preference scores and the event preference scores to obtain user event bidirectional preference scores; and secondly, linearly weighted combining the topic matching degree score and the bidirectional preference score to obtain a final recommendation degree score of the user-event pair. The following is a specific description.

1) The user-event pairs are scored for bi-directional preference. Assume that the preference scoring weights for the user and event are respectivelyAndthe user event bidirectional preference score +.>. The key issue of bi-directional preference scoring is then to find the weight vector of the two preference scores, choosing to learn the weight vector using implicit feedback as training data. Unlike explicit feedback where the user scores items, implicit feedback in an event social network can only be represented by the interaction information between the user and the event, i.e., if the user attended the event, the feedback is 1, otherwise the feedback is 0. Obviously, the user's feedback is 0 for all new events.

The learning algorithm BPR based on Bayesian maximum likelihood estimation is selected to perform rank learning on the weights, and the correct rank order of the user-event pairs is learned according to implicit feedback data of the user on the events, so that the events participated by the user are ranked before new events or other events. First, define the maximum posterior probabilityAs shown in formula (26).

(26)

wherein ,θthe weight vector is represented by a weight vector,Rrepresenting a set of all user-event pairs,the definition is shown in formula (27).

(27)

wherein ,representing user +.>User-event pairs of->Representing +.>Event->Arranged at->The former probability is shown in equation (28).

(28)

wherein ,i.e. bi-directional preference score->，. For more convenient optimization, assume +.>Obeying the normal distribution with the mean value of 0, developing and deducing to obtain a final optimized objective function +.>As shown in formula (29).

(29)

wherein ,representing the regularized term coefficients. And (5) maximizing and optimizing an objective function through implicit interactive feedback data of the user event, and obtaining an optimal weight parameter vector. Solving the optimization problem using a random gradient descent algorithm (Stochastic Gradient Descent, SGD), randomly extracting user-event pairs of the target user from the training set in an iterative process to update the weight vector +. >The update process is shown in equation (30).

(30)

wherein ,is learning rate (I/O)>. Through the learning process, the training set and the super parameter ++can be automatically scored according to the user event preference> andObtaining a weight vector->Thereby obtaining a bi-directional preference score->。

2) And obtaining the final recommendation scores of the user-event pairs by combining the topic matching and the bidirectional preference. Combining the discussion about topic matching and preference calculation of users and events, firstly, extracting event topics through an LDA topic model and obtaining topic matching degree scores of the users and the events; secondly, respectively constructing preference models of users and events according to user event context information in the EBSN, and obtaining bidirectional preference scores of the user events through a BPR learning algorithm; finally, scoring the topic matching degreeBi-directional preference score with user event->Linear weighted summation to get final user-event pair recommendation score +.>As shown in formula (31). />

(31)

wherein ,for weight parameters, typically manually set empirically, the optimal setting will be determined experimentally. Algorithm 4 describes a process of fusing topic matching and bi-directional preference solving for user-event scores for final recommendation.

Algorithm 4 gives the final resultA process of fusing topic match scores and user event bi-directional preference scores. Firstly, training sets generated by user preference score sets and event preference score sets are subjected to sequencing learning through a Bayesian personalized sequencing algorithm to obtain an optimal weight vector And according to->Calculating user-event pair bidirectional preference scores for the target user (line 2 to line 10); secondly, the topic matching degree scores and the bidirectional preference scores of the user-event pairs are linearly combined to obtain final recommendation degree scores (11 th line to 13 th line), so that TOP-K events are recommended to the user according to the final recommendation degree score ranking.

So far, we have combined topic matching and user event bi-directional preferences, put forward a personalized event recommendation scheme, and detailed the specific content of which is described in the above section.

In the personalized event recommendation method and system based on fusion of the topic matching and the bidirectional preference, firstly, the topic information of the event is extracted by utilizing the document topic generation model LDA, the user topic information is obtained according to the historical event record of the user participation, the topic matching degree of the user and the event is calculated to be used as an important recommendation factor in the recommendation model, and the topic factor can better represent the feature preference; secondly, considering the social network recommendation based on the event from the two-way angles of the user and the event, constructing preference models of the user and the event, respectively obtaining user preference scores and event preference scores, and more completely mining preference relations from the two angles of the user and the event; finally, the user-event pair matching degree is combined with the user event bi-directional preference linear weighted combination to obtain the final user-event pair comprehensive score, and the ordered TOP-K user-event pairs are used as recommendation results. A large number of experiments are carried out on the Meetup real data set, and the comparison is carried out with other event recommendation algorithms, so that the performance of the software recommendation algorithm is superior to that of the traditional recommendation scheme, the personalized preference of a user can be well predicted, and the purpose of personalized recommendation is achieved.

It should be noted that the above-mentioned embodiments are merely preferred embodiments of the present invention, and are not intended to limit the present invention, but various modifications and variations of the present invention will be apparent to those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. The personalized event recommendation method integrating theme matching and bidirectional preference is characterized by comprising the following steps of:

learning weight parameters of user preference scores and event preference scores by using a Bayesian personalized ranking algorithm (BPR), obtaining user event bidirectional preference scores, linearly weighting and combining the topic matching degree scores and the bidirectional preference scores to obtain final recommendation scores of user-event pairs, and recommending the top K events after ranking to the user;

The calculating the theme of the new event and the user history event in the first step adopts a JS divergence algorithm to calculate the theme matching degree score of the user-event pair, and the specific steps include:

(1)

given set of documents， andRespectively represent document->Is a priori Dirichlet distribution of topic distribution and word distribution,/->The superparameter of the subject prior distribution and the word prior distribution given empirically,kis the number of topics of the previously specified document set,N _m representation document->Is used to determine the number of words in the word,Mis the number of documents in the document set; for documents->LDA is based on a priori knowledge +.>Determining topic distribution of a document>Then distribute->Extracting a themezYet according to a priori knowledge->Determining word distribution of a current topic- >From the subjectzCorresponding word distribution->Extracting a word->Repeating the above processN _m The document +.>；

the topic distribution of all events has been generated according to equation (1)Given event-> andRespectively have theme distribution->The JS divergence +.>As shown in formula (2):

(2)

wherein ,，represent KL divergence to describe two probability distributions +.> andThe difference between the two is calculated as shown in formula (3):

(3)

(4)

(5)

2. The personalized event recommendation method integrating topic matching and bi-directional preference according to claim 1, wherein the document topic generation model LDA in step one has a three-layer generation type bayesian network structure comprising documents, topics and words, wherein both document-topic and topic-word obey polynomial distribution; each document selects a topic with a certain probability, and a word is selected from the topic with a certain probability, and topics in any document conform to the Dirichlet distribution, and relationships among texts are discovered through the distribution.

3. The personalized event recommendation method of integrating topic matching and bidirectional preferences according to claim 2, wherein the constructing a user preference model in the second step constructs single-factor preferences of a user from three aspects of geographic location, social relationship and time factor, respectively, specifically comprising:

step 2-1-1, constructing a geographic location preference model:

the geographic position preference model calculates the probability that a target user will participate in holding an event at the position, a kernel density estimation KDE method is adopted to model the two-dimensional geographic position distribution of the event in which the user participates, and the event participation probability after normalization is used to represent the preference degree of the user to the geographic position; longitude and latitude coordinates of event geographic position Lx, Ly) Representing a set of places where a user historically participated in an eventL(u) Representation, then about the useruKDE function of (F)As shown in formula (6):

(6)

wherein ,l _i =(Lx _i , Ly _i ) ^T a two-dimensional vector representing the longitude and latitude coordinates of the event location,m _l (u,l _i ) Representing a useruParticipating geographic locationsl _i The frequency of the holding-up of the event is at the process,σrepresenting the size of the neighborhood window (bandwidth),Nthe number in the position sample is indicated,K(•) The Gaussian kernel function is represented, and the definition form is shown as a formula (7):

(7)

(8)

(9)

step 2-1-2, constructing a social relation preference model:

first, the relevance of users to a group is defined as the relevance of users to the place they belong to With interactive relationships between groups and between users and events created within a group toG(u) Representing a useruSet of groups to which the participating events belong, then the relevance of the user to the groupCan be represented by the formula (10):

(10)

(11)

(12)

will bes(u,g) Normalized toAs shown in formula (13):

(13)

in combination with the above-mentioned two kinds of interaction relations,users belonging to the same group tend to participate in events created by other users within the group, and the user is derived by integrating the user's relevance to the group and the intra-group user relevanceuWith respect to on-line teamsgSocial preference scoring of (2)As shown in formula (14):

(14)

step 2-1-3, constructing a time factor preference model:

The time factor of the event is an important preference factor to be considered when calculating user preferences; new event with optional participation of usereRepresented as a 7 x 24 dimensional event time vectorWhen a new event occurs in a certain specific time period of the week, setting the vector component value of the time period to be 1, otherwise, setting the vector component value to be 0; representing the user as a user time vector +.>As shown in formula (15):

(15)

(16)

(17)。

4. the personalized event recommendation method combining topic matching and bi-directional preferences according to claim 3, wherein the calculating a user preference score in step two specifically comprises:

for the geographic position preference model, representing a geographic position preference score by predicting the probability of a user participating in event activities held at the position; calculating social preference scores of the target users from two aspects of the relationship between the target users and the group and the relevance between the target users and the users in the group for the social relationship preference model; for the time factor preference model, by constructing two granularities of date and hour Unifying the vector representations and calculating the similarity of the user-event pairs as a time preference score for the target user based thereon; combining the three single-factor preferences to form a user preference perception model, and obtaining the user by linearly combining the three single-factor preferencesuFor eventseOverall preference scoring of (a)As shown in formula (18):

(18)

5. The personalized event recommendation method according to claim 4, wherein the constructing an event preference model in the second step constructs single factor preferences of the event from two aspects of event location popularity and event sponsor influence, respectively, specifically comprising:

step 2-2-1, constructing an event location popularity preference model:

first define the geographic location of an eventl _e In relation to usersuPopularity of (2)As shown in formula (19):

(19)

wherein the moleculeFor usersuParticipating geographic locationsl _e Frequency of holding activities is treated as useruThe maximum frequency of historic visited locations; likewise, a geographic location is defined l _e In relation to usersuGroup ofgPopularity of->As shown in formula (20):

(20)

(21)

step 2-2-2, constructing an event sponsor influence preference model:

(22)

(23)

wherein ,U _g representation teamUser set in- >Representing user +.>Participating sponsors->Event collection held,/->Representation->In the group->A set of events held in the middle; obtaining comprehensive influence degree scores of event sponsors by combining influence degree of the event sponsors on target users and on users in the group>As shown in formula (24):

(24)。

6. the personalized event recommendation method combining topic matching and bi-directional preferences according to claim 5, wherein the calculating the event preference score in step two specifically comprises:

(25)。

7. the personalized event recommendation method for merging topic matching and bi-directional preferences according to claim 6, wherein the obtaining a user event bi-directional preference score in step three, linearly weighted combining the topic matching degree score and the bi-directional preference score to obtain a final recommendation score for the user-event pair, comprises the specific steps of:

step 3-1, two-way preference is made for user-event pairs:

selecting a learning algorithm BPR based on Bayesian maximum likelihood estimation to perform sequencing learning on weights, and learning a correct sequencing order of user-event pairs according to implicit feedback data of the user on the events so that the events participated by the user are ranked before new events or other events; first, define the maximum posterior probabilityAs shown in formula (26):

(26)

(27)

(28)

(29)

wherein ,representing the regular term coefficient, and obtaining an optimal weight parameter vector by using implicit interactive feedback data maximization optimization objective function of a user event; solving the optimization problem by adopting a random gradient descent algorithm SGD, randomly extracting user-event pairs of target users from a training set in an iterative process to update weight vectors +. >The update process is as shown in equation (30):

(30)

(31)

8. An implementation system for integrating topic matching with bi-directional preference personalized event recommendation, which is used for implementing the method for integrating topic matching with bi-directional preference personalized event recommendation according to any one of claims 1-7, wherein the implementation system comprises:

9. The system for implementing personalized event recommendation that merges topic matching with bi-directional preferences of claim 8, wherein the user preference module comprises a geographic location preference module, a social relationship preference module, and a time factor preference module, the event preference module comprises an event location popularity preference module and an event sponsor influence preference module, wherein: