Location Recommendation And Its Application In Location-based Social Network

The pervasive use of Location-based Social Network (LBSN) makes location rec-ommendation one of the most prevalent applications. Location recommendation aims to recommend locations of users’ favorite, in which location refers to merchants, orga-nizations and public spots in urban area. Taking advantage of location recommender system, users can explore the city, find fascinating merchants or places that are of their interest thus getting enriched experience. Besides, merchants can benefit as well by increasing their revenue through immediate virtual marketing.There is an abundance of polygenetic, heterogeneous information in LBSN, such as social correlation, geographical influence, user’s textual reviews or ratings and temporal effect. Undoubtedly, we can make more precise and personalized recommendation by utilizing these aforementioned information. However, there are still several challenges to come along. Firstly, similar to conventional recommender systems, cold start prob-lem due to data sparsity is needed to be tackled with. Secondly, it is another question worthy of research that how to fuse all the polygenetic and heterogeneous information into one model. In order to address all the aforementioned problem, this dissertation aims to delve into location recommendation and its application in LBSN, that is to u-tilize as much information in LBSN as possible to model users’ visiting behavior, thus predicting users’ preferences in order to make recommendations. To be specific, our works, conclusions and along with the contributions are as follows:1)Circle-based Social Correlation Mining for Location Recommendation. User-s in social networks tends to have similar preferences with their neighbors due to the "homophily phenomenon", which provides the theory basis for learning user prefer-ences. Accordingly, we model users checkin behavior in each interest circle which was created based on the categories of visited POIs. We devise two regularization terms in accordance with the friends’ and experts’ influence respectively, thus resulting in a circle-based and social-regularized model in order to learn users’ preferences. Experi-ments on a real world dataset show the superiority of our approach, especially in solving cold-start problem.2)Learning User’s Comparative Choice for Location Recommendation. A user’s rating behaviors are not independent with each other, which means a user’s rating is made in terms of her historical visits. The rating is finally given under rational compar-ison with her historical experiences of the ratings. By leveraging the relativity, compar-ison and recency properties, we learn users’ recency based comparative choice towards location recommendation. Choice model which is derived from utility theory in the economics is employed to build the optimization function and meanwhile a collection-wise stochastic gradient descent algorithm is devised to learn users’ preferences. We evaluate our method on two large-scale datasets and the result shows that our approach outperforms the state-of-the-art methods.3)Mining User’s Multiple Preferences for Location Recommendation. Different users may have different preferences, even for those who give the same rating values to the same place, which is because users preferences across multiple aspects follows distinct distributions. Upon the assumption, we mine the textual reviews to obtain a user’s explicit interest and utilize the probabilistic latent factor to depict her implicit interest, and thereby forming a unified function to fit the rating data. A gradient descent algorithm is devised for the loss function thus the users preferences are obatined. We not only conduct extensive experiments on two real-world datasets which indicates that our approach is superior to other methods and meanwhile solving the cold start problem, but also implement a mobile APP in order to make real-time recommendation for online users.4)Exploring polygentic and heterogenous information for Location Recommenda-tion. Location-based Social Network contains numerous information including social network, geographical location, textual review and so on. In this work, we propose a probabilistic latent factor model based on the aforementioned information. We adopt k-ernel density estimation and friend-based collaborative filtering to model the geographi-cal influence and social correlation respectively. Topic model is utilized to extract users’ explicit interest while users’ implicit interest is depicted by latent factor model. Exper-iments on a real world dataset exposit our approach with multi-sources outperforms the others. Moreover, our model show its robustness due to its modularization.