Research On Group Recommendation Models Based On Latent Representation For Event-based Social Networks

In recent years,various social networks emerge in endlessly with the continuous development of mobile Internet.The means and methods of people’s mutual communication have changed greatly.The combination of online and offline interaction has become a new development trend.The event-based social network(EBSN)is a social platform that can realize the combination of online and offline.At present,it has become more and more popular.For example,there are Meetup,Plancast and Douban,etc.On such platforms,organizers can publish events online,while users can query the relevant information of events and participate in different real events offline.So it enhances mutual understanding between users.However,with the continuous development of EBSN,the problem of information overload has become increasingly prominent.It is difficult for users to find the events they are interested in.The search engine and recommendation system are two filtering technologies that can effectively alleviate the problem of information overload.Different from search engine,the recommendation system doesn’t require users to enter keywords for query.It is an implicit and active information filtering technology.The recommendation system can portray users according to their access records and scoring information,so as to provide users with goods they may be interested in.Therefore,the event recommendation system becomes an important means to solve the problem of information overload in EBSN.The traditional event recommendation system usually recommends events to a single user.However,users in EBSN usually join a group and participate in various events as group members.For example,users watch basketball games with friends,participate in academic activities with colleagues,watch movies with family members,and so on.Therefore,it is very meaningful and necessary to carry out group event recommendation model algorithm research in EBSN.As we all know,the recommendation algorithm based on collaborative filtering is one of the most classical algorithms in the recommendation system.The latent representation model is the most widely used class of collaborative filtering recommendation algorithms.Specifically,the idea of latent representation model is based on the scoring matrix between users and items,which maps users and items into an latent vector space to obtain the latent representations of users and items respectively.Then the latent representations are used to model the user’s score,so as to predict the user’s score on the projects.The latent representation model is easy to be implemented and has high prediction accuracy.It is widely used in various recommendation systems,including group recommendation system in EBSN.However,the recommendation effect is still poor in EBSN if only the score matrix composed of groups and events is used for recommendation.The main reason is that this type of matrix is extremely sparse.At the same time,through literature review and in-depth analysis of the internal principle of latent representation model,we found that there are still the following main challenges in the application of traditional latent representation model used in EBSN group recommendation: 1)The group’s decision to go offline to participate in an event is affected by fine-grained explicit feature factors and the different explicit feature factors have different degrees of influence.The traditional latent representation model considers groups and events as an independent whole from the perspective of coarse granularity.On the one hand,the model lacks good explanatory ability.On the other hand,the group preference vector representation is fixed,resulting in the lack of "personalization" of the learned group vector implicit representation,which affects the recommendation accuracy.2)Besides users and groups having rating data for events,the EBSN also contains rich implicit social information.This part of information can assist in recommendation and alleviate the problem of data sparsity.The group event recommendation method based on the traditional latent representation model lacks the mining and application of this part of the data.3)There may be a certain correlation between explicit information and implicit information in EBSN.When the traditional latent representation-based group event recommendation methods use explicit and implicit information to model,they usually model the two separately.It leads to redundancy in the learned implicit feature vector representation,which reduces the recommendation accuracy.4)The general latent representation-based group recommendation methods usually assume that the preferences of users in EBSN are static and invariant.They lack consideration of the dynamic variability of user preferences.At the same time,they don’t learn higher-order implicit information when using the implicit social relationship.In response to the above four challenges,this thesis is based on the traditional latent representation model and integrates a large amount of explicit information and implicit information in EBSN.In this study,considering the dynamic preferences of users and groups,combining with the idea of deep learning,representation learning technology and hybrid recommendation,the four novel group recommendation methods based on hybrid deep latent representation are proposed to provide more accurate recommendation service for groups in EBSN.The main contents and innovations of this thesis are summarized as follows:(1)Aiming at the problem that the vector representation learned by the traditional latent representation model is fixed that leads to the degradation of recommendation performance,and the model lacks good explanation,this thesis proposes a hierarchical attention latent representation group recommendation model HALR based on explicit features.Specifically,the decision-making of group participation events in EBSN is affected by their respective explicit feature factors and the influence degree of different explicit features is different.We use a variety of explicit features and their different weights to express groups and events.Firstly,the explicit features of groups and events are mapped to low dimensional dense semantic vector space through embedded layer neural network.The topic features of groups and events are extracted by convolutional neural networks(CNN)and attention mechanism.Then,the attention neural network is used to learn the influence weights of the extracted topic features and other explicit features.The explicit features representations are weighted and summed.The high-level implicit vector representation of groups and events are obtained.Finally,the group score prediction is modeled based on the implicit vector representation of groups and events.Experiments are carried out on five real EBSN data sets.The results show that the vector latent representation learned by HALR model is more personalized.The model has good interpretability and significantly improves the recommendation performance compared with the existing methods.Compared with the optimal model method in the baseline,the RMSE performance index of HALR model is improved by an average of 3.35%and the MAE performance index of HALR model is improved by an average of 4.32%.(2)Aiming at the problem that the traditional latent representation model has sparse data,which affects the recommendation effect,this thesis proposes a graph attention latent representation group recommendation model SRGAM based on implicit social relationship.Specifically,there are valuable implicit social relationship information in EBSN.This kind of information can assist recommendation and alleviate the problem of data sparsity.The SRGAM model not only constructs user-event interaction graph and event-user interaction graph,but also builds user-user social graph and event-event social graph.It further alleviates the problem of data sparsity in group event recommendation.We use graph attention neural network to learn graph data and calculate the different influence weights between nodes in heterogeneous graphs,so as to produce more reasonable latent vector representations of users and events.Then,we further use the attention mechanism to weighted fuse the latent vector representations of different users in the group to obtain a high-level group latent vector representation.Finally,we model the group score prediction based on the latent vector representation of groups and events.We selected five data set on the Meetup website for experiment and compared the SRGAM model with a variety of advanced methods.The results show that the SRGAM model alleviates the data sparsity and makes the recommendation performance better.Compared with the optimal model method in the baseline,this model has an average increase of 5.34% in RMSE performance index and4.81% in MAE performance index.(3)Aiming at the problem that the traditional latent representation model models separate the explicit information and implicit information in EBSN,which leads to the redundancy of learning vectors and the poor performance of the model,this thesis attempts to model them uniformly and proposes a graph multi-head attention implicit representation group recommendation model GMAN combining explicit and implicit information.Specifically,the GMAN model aggregates groups and events with explicit information and implicit information respectively.When aggregating,we take the vector representation of the explicit aggregation output as the input of implicit information aggregation.We construct a unified model of user’s and event’s explicit information and implicit information by using graph neural network.Thus,the problem of data sparsity is further alleviated and the problem of learning vector redundancy is solved.In addition,we use multi-attention mechanisms to learn a richer vector representation of users and events from many different perspectives.Moreover,we use the native attention mechanism to aggregate the user vectors in the group and finally generate high-level group feature vectors.Experiments on real data sets show that the performance of our GMAN model is better.Compared with the optimal model in the baseline,this model has an average increase of 15.94% in RMSE performance index and 9.43% in MAE performance index.(4)Aiming at the problems that the traditional latent representation model fails to consider the dynamic variation of user preferences and the influence of higher-order structural information in implicit social relationships,which still has room to improve the performance of the model,this thesis proposes an latent representation group recommendation model LSHAG based on long-term and short-term preference heterogeneous attribute graph.Specifically,we take the recent interactions between users and events as the user’s short-term preference,and take the complete historical interactions between users and events expect the recent interactions as the user’s long-term preference.Because users’ high-order friends(such as friends of friends)may affect users’ preferences,this thesis further uses the high-order structure information in social relationship to assist recommendation.Then we use graph neural network to encode explicit information,implicit information and higher-order structural information in EBSN,and propagate the embedded vector on the graph so as to learn the latent vector representation of users and events.Finally,the multi-layer perceptron and attention mechanism are used to fuse the user’s long-and short-term preferences and generate latent vector representation of the final group’s preferences.Extensive experiments on several real data sets show that our LSHAG model performs better than the existing advanced models.Compared with the optimal model in the baseline,this model has an average increase of 5.76% in Pre@5 performance index and 13.92% in Rec@5 performance index.