Research On Exploiting Multi-dimensional Behavior Characteristics For Point-of-interest Recommendation Models

In real life,users often visit locations in the real world and achieve their relevant purposes by visiting these locations.Thanks to the rapid development of location-based Social Networks(LBSNs)in recent years,LBSNs provide users with a platform to share life and experience.Users record and share the real-world location they visit to form user check-in behaviors.These locations where users check-in can be called Point-of-Interest(POI).However,as more and more POIs are included,a large number of POIs cause the problem of"information overload".The emergence of personalized POI recommendation system solves the above problems by integrating various contexts to recommend POIs that users may interest.Although personalized POI recommendation system has been widely studied,existing studies lack in-depth mining of the phenomena contained in check-in data,user behavior characteristics and the relationship between characteristics,so as to improve the accuracy of recommendation and alleviate the problem of data sparsity in POI recommendation.The main contributions and innovations of this thesis are as follows:(1)Research on the POI recommendation based on multi-dimensional characteristics of category level.Firstly,it is found that the new phenomenon reflected in the user’s check-in frequencies and the Diminishing Marginal Utility contained in check-in frequency characteristics.It is found that the user’s visit to different POI categories will be influenced by the user’s historical check-in frequency characteristics of these POI categories.A more accurate probability adjustment mechanism for users to visit specific POI categories is formed by improving KL-divergence and a text analysis algorithm.Secondly,a categorylevel sequential-and non-sequential influence-aware probabilistic generative model is proposed to model the relationship between the sequence characteristics of user check-in behavior and users’preference for the POI category.Finally,the above behavior characteristics at the category level are integrated into a unified method.The experimental results on two real check-in datasets are significantly better than the existing methods.(2)Research on POI recommendation integrating the user internal and external characteristic.Firstly,through data analysis,it is found that there is a Status Effect in the check-in behavior of users in their corresponding community at the category level,and the effect is expressed and modeled through community influence power.Secondly,a new user community detection algorithm is proposed,which generates user communities from the two dimensions of user behavior and activity area with the help of word embedding technology and topic model.Then,according to the generated user community,a user community influence power-aware generative model is proposed.The proportion of users’ real-time check-in frequencies and the number of real-time check-in POI categories are jointly mapped to the model’ s user community influence power module.Finally,the proposed model models user community influence power characteristic(external characteristic)and user preferences(internal characteristic)at the same time,and explains the relevant conditions of their respective effects.The experimental results on two real check-in datasets are significantly better than the baselines.(3)Research on POI recommendation integrating user’s one-time check-in characteristic.First of all,the check-in behavior is defined,and through data analysis,it is found that one-time check-in behavior accounts for a considerable proportion in users’ check-in behavior,and one-time check-in behavior is understood with the help of Involuntary Attention Theory in psychology,and it is found that the relationship between one-time check-in behavior and category,as well as the relationship between one-time check-in behavior and moving distance range.Then,a user one-time check-in characteristic-aware probability generative model is proposed.In this model,the user preference and one-time check-in pattern are modeled,and the relationship between them is modeled at the same time.In addition,the characteristics of moving distance range and the check-in frequencies at the POI level are used to further alleviate the problem of data sparsity,and help to explain the generation conditions of one-time check-in behavior.The experimental results on two real check-in datasets are significantly better than the baselines.(4)Research on POI recommendation based on the user functional requirement characteristic.Through data analysis,it is found that there are obvious differences in the check-in frequencies of users’ POI in the same category.It is considered that this difference is caused by users’ specific latent functional requirement and this difference is attributed to the Halo Effect in psychology.Based on this discovery,a user functional requirement characteristicaware topic model is proposed,which adaptively models user preference and user functional requirement characteristic,and divides user behavior into user preference dimension and functional requirement dimension,so that the model can adapt to the characteristics of user preference dimension and functional requirement dimension.The model adds the moving distance range characteristic in the user preference dimension and adds the factor of check-in frequencies of POIs in the functional requirement dimension.In addition,the check-in frequency characteristics at the category level and the POI level are integrated into the process of generating recommendation,which further alleviates the problem of data sparsity and optimizes the recommendation results.Experiments on two real check-in datasets show that this method has significant advantages over the existing methods.