Personalized Point-of-Interest Recommendation Via Deep Neural Networks

With the development of mobile Internet technology,location-based social networks(LBSNs)have become an important part of people's lives and brought a large amount of user check-in data.Personalized point-of-interest(POI)recommendation is an important application in LBSNs which utilizes users' check-ins to learn their preferences for POls and provide personalized recommendation services.However,personalized POI recommendation usually suffers from the data sparsity issue due to the limited check-ins of a user.Furthermore,how to model the spatio-temporal features in LBSNs and learn users' dynamic preferences is also very important for personalized POI recommendation.In light of these issues,this paper has developed a novel personalized POI recommendation framework which is based on deep neural networks(DNNs).The framework employs a symmetric matrix factorization(MF)to obtain vector representations of POIs' attributes.Then,users,POIs,temporal and sequential contexts are represented as in the form of feature vectors according to users' check-in histories.To predict users' personalized preferences for POIs in the given spatio-temporal contexts,a DNN-based recommendation model is designed,which takes feature vectors of users,POIs,temporal and sequential contexts as inputs.To summarize,our works and contributions are as follows.1)Users' check-in behaviors are analyzed in this paper based on two public LBSNs check-in datasets.To be specific,geographical,temporal and sequential influences on users' check-ins have been investigated.The results have shown that users'check-in behaviors are strongly influenced by the spatio-temporal contexts of their check-ins and motived us to design a novel personalized POI recommendation framework to model these influences.2)A novel MF-based feature representation method is developed to utilize various features in LBSNs to alleviate the data sparsity issue and model the spatio-temporal features.Firstly,we have constructed a co-occurrence matrix,a geographical proximity matrix and a categorical correlation matrix based on users' check-ins and POIs' attributes.Then,a symmetric matrix factorization is applied on the three feature matrix to obtain low-dimensional vector representations of POIs' features.At last,users,POIs,temporal and sequential contexts are represented as feature vectors based on users' check-in records.3)A DNN-based personalized recommendation model is proposed to model non-linear feature interactions and learn users' dynamic user preferences.The model addresses the personalized POI recommendation problem as a binary classification task and takes feature vectors of users,POIs,temporal and sequential contexts as inputs.Then,a DNN is employed to learn a user's preferences for POIs in the given contexts and give the predicted scores.Top-k POIs with highest scores are presented to the corresponding user as recommendation results.4)Extensive experiments are conducted on the two check-in datasets to demonstrate the effectiveness of our proposed framework.To be specific,we have compared our method with state-of-the-art methods for personalized POI recommendation.Experimental results on both datasets have shown that our method outperforms other comparative methods significantly in both traditional and context-aware scenarios.Then,we have design several simplified frameworks to show the contributions of different factors in our framework.The impacts of network size have been investigated by our paper as well.