Research On User Activity Patterns-aware Mobile Recommendation Approaches In Location-based Social Networks

With the development of mobile Internet,location-based social networks(LBSNs for short),such as foursquare and yelp,have been widely developed.Users share their experiences on location-based social networks,such as sharing delicious food,scenic spot photography,etc.,and form friends with other users online.However,the development of LBSNs has also brought serious "information overload" problem to users,because it is difficult for users in LBSNs to find attractive items from a large number of points of interest(POI for short).Fortunately,POI recommendation systems have become an effective means to alleviate the problem of information overload in LBSNs.Compared with traditional recommendation,POI recommendation has a more serious problem of data sparsity.The existing methods make full use of abundant relevant data in LBSNs to alleviate the data sparsity problem.However,these methods ignore the fact that users choose the corresponding POIs in order to participate in activities and lack in-depth research on the complex activity patterns of users,so they can not distinguish the dynamic impact of various factors on user decisionmaking under the activity patterns,which results in that the user check-ins formed by the influence of different factors are used for modeling user personal preferences.Furthermore,it is difficult to accurately depict user preferences,which brings a great challenge to the POI recommendation.Based on the above challenge,this paper investigates POI recommendation based on user activity pattern perception in location-based social networks by analyzing user behavior patterns from the perspective of activities,and then accurately mine user preferences under activity patterns,further provide accurate and efficient point of interest recommendation services for LBSNs' users.The main contents and innovations of this paper are summarized as follows:(1)A POI recommendation based on user activity time pattern perception is proposed to effectively model complex user activity time patterns.Firstly,throught data analysis,it is found that users' personalized preference is not a single factor affecting their time decision,and users have complex activity time patterns in the real environment.On this basis,the user preference is divided into user personalized activity interest and their social preference,in which the latter reflects a user's tendency in the choice of POIs under the influence of activities.And then the user activity time pattern is dividied into activity time pattern under the influence of his personalized activity interest,activity time pattern under the periodic influence of his social preference and activity time pattern under the aperioidc effect of his social preference.In addition,Alias sampling method is used to improve the training efficiency of the model.The experimental results on two real location-based social network datasets demonstrate that the POI recommendation method based on user activity time pattern perception has significantly improved the recommendation accuracy and model training efficiency compared with the existing methods.(2)A POI recommendation method based on user activity spatio-temporal pattern perception is proposed,which effectively models the diversity of user activity patterns in spatio-temporal context.Firstly,user activities are mapped into the newly defined spatio-temporal activity map,which reveals the diversity of user activity patterns in spatio-temporal contexts and exhibits the obvious differences of user decision-making in the same activity pattern.On this basis,a probabilistic generation model is proposed to jointly model user activity patterns and their preferences,so as to accurately describe user spatio-temporal activity patterns and precisely capture the user preferences under the complex user spatio-temporal activity patterns.Experimental results on two real LBSNs datasets show that the proposed method outperforms the existing methods.(3)A POI recommendation method based on user check-in pattern perception under the influence of spatio-temporal activities.Firstly,through data analysis,it is found that there are three phenomena related to activities in the process of user check-ins,that is,the social effect of activities,the inconsistency between users' social preferences and decisions,and the correlation between the characteristics of activity frequency and the key factors affecting check-in.On this basis,a probabilistic generation model is proposed,which further subdivides user preferences into personal activity preferences and their social preferences,and establishes a connection between the latent topics representing the two preferences,and excavates the social impact of user activities.Combined with the activity frequency characteristics,a joint switch is designed.According to the correlation,the check-ins are dividied into two categories,namely the check-ins affected by user social preferences and the check-ins affected by spatio-temporal crowd.The influence weight of user social preferences in their decision-making is adaptively learned from the check-in samples to obtain the complex user check-in patterns under the influence of spatio-temporal activities and accurately depict the user preferences.The experimental results on two real location-based social network datasets illustrate that the proposed method has significantly improved performance compared with the existing methods,and the training efficiency is also higher than the traditional method based on Gibbs sampling,and the multivariate Gaussian distribution can collect more coherent latent topics.(4)A POI recommendation method based on user activity sequence patterns perception is proposed to alleviate the newly discovered cold-start sequence problem.Firstly,an activity sequence semantic representation strategy is proposed.Combined with the semantics of activities in the sequence and the corresponding time interval information,an accurate activity sequence representation vector is generated,and then the similarity of user check-ins under the activity sequence with similar semantics is found.Based on this phenomenon,a probabilistic generation model is proposed,which maps different,sparse and semantically related activity sequences to similar positions in semantic space,and uses multivariate Gaussian distribution to extract semantically continuous sequence patterns.Afterward,it adaptively learns the influence weights of users' social preferences and spatio-temporal crowd effects on their decision-making under activity sequence patterns.Experimental results on two real LBSN datasets show that the proposed method has higher recommendation accuracy than the existing methods.In addition,the proposed model can obtain user activity sequence patterns more flexibly.