Research On Meta-path And Multi-factor Fusion Based POI Recommendation And Successive POI Recommendation

In recent years,with the development of global positioning system and the universal use of location aware mobile devices,location-based social networks(LBSN)have developed rapidly and become popular.As an important service in LBSN,Point of Interests(POI)recommendation has attracted extensive attention from academia and industry.POI recommendation recommends POI to users according to users' personal preferences by analyzing their check-in records.However,compared with the traditional item recommendation such as Taobao,POI recommendation in LBSN faces a more serious problem of data sparsity.Among millions of POIs,a user can only access a small part of POIs.In addition,existing research on POI recommendation is often based on homogeneous information networks,ignoring the different types of nodes in LBSN and the different relationships between nodes,which makes it difficult to integrate various heterogeneous information.Moreover,most of existing POI recommendation methods only consider few influential factors,and these methods are difficult to extend to include all the factors in a unified way.As a natural extended task of POI recommendation,successive POI recommendation began to attract people's attention.The purpose of successive POI recommendation is to recommend POI to user at successive time.Compared with POI recommendation,successive POI recommendation not only needs to consider the user's personal preference,but also attach more importance to the sequential information of users' check-in behavior.However,the user's sequential behavior patterns are complex and changeable,it is still a challenge to model the user's sequential behavior patterns combined with spatio-temporal factors.In addition,users' successive check-in behavior is impacted by dynamic social influence,and it is difficult to identify the correlation between the check-in behavior of users and friends.This paper focuses on recommendation tasks in LBSN and the main research contents and innovations are summarized as follows:(1)Due to defects of existing POI recommendation methods,this paper proposes a meta-path-based deep representation learning method(MPDRL)for personalized point of interest recommendation.We first propose a meta-path similarity matrix based representation method for POIs,which incorporates the geographical,temporal,categorical,and co-visiting information in a heterogeneous information network.In order to comprehensively consider the influential factors that affect the user's check-in behavior,we further design an attention based long short-term memory network to mine the complex correlations between users and POIs.(2)In order to further solve the problems in successive POI recommendation,this paper proposes a sequential behavior and dynamic social influence modeling(SBDS)method for successive POI recommendation.We first propose a spatio-temporal based attention mechanism to capture the importance of different POI with considering the temporal context and obtain representation of users' long-term behavior sequence.Then,we designed a convolutional neural network with different convolution kernels to model the discontinuity and complex sequential patterns of the user's short-term behavior sequence.The final behavior sequence representation of users is obtained by considering the different dependence of different users on long-term interest and shortterm intention.Moreover,this paper proposes a representation method for dynamically modeling users' social influence based on different contexts.Finally,the sequential information and social influence are integrated to recommend successive POI for users.(3)Experimental results show that the proposed POI recommendation model MPDRL has shown at least 16.97% and 23.55% improvement in terms of the metric precision and recall compared with the current popular model.Our proposed successive POI recommendation model SBDS improves at least 11.94% and 16.62% in terms of the metric recall and mean average precision compared with the current popular model.