| With the rise of social media based on web 2.0 technology, the emergence of location-based social networking services and many kinds of mobile terminal social media, as well as the rapid development of the city, the number of POIs are also increasing,people are willing to explore the cities and the neighboring places in their daily life,according to their personal interests with regard to the kinds of choices of POIs to decide where to go. Location-based social networks provide an unprecedented opportunity to study human mobile behavior for POI recommendation, users like to share their check-ins of various places and interests, as well as their experience for products, services and comments, build and maintain their social relationship in these LBSNs platforms to show their interest and personality. The founders of these location-based social networks also pay more attention to collect and analyze essential data and behavior data of users,to better understand user mobile behavior,to know more about their users, exploit POI recommendation to improve the users' experience and meet the needs of users. At the same time POI recommendation in social media faces some new problems. How to comprehensive use of various data in social media, how to solve the sparsity problem of users' check-in data, how to deal with the implicit users' feedback and complex users' relationship, how to deal with the timeliness of user generated content, these problems face with challenges. According to these challenges, this paper puts forward and designs a series of context-aware user model and POI recommendation algorithm, improves the effect of POI recommendation and the users' experience in social media. The innovations and contributions of this paper are summarized as follows:1. Context-aware POI recommendation in location-based social networksThe user visited POI occupied very small proportion in location based social network, the use-POI check-in matrix is extreme sparse, the interest of user is the dynamic change at different time and geographic location. To solve these problems, this paper proposes a context-aware probabilistic matrix factorization method for POI recommendation. First, this paper exploits an aggregated Latent Dirichlet Allocation (LDA) model to learn the interest topics of users and infers the interest POIs by mining textual information associated with POIs and generates interest relevance score. Second, this paper proposes a kernel estimation method with an adaptive bandwidth to model the geographical correlations and generates geographical relevance score. Third,this paper builds social relevance through the power-law distribution of user social relations to generate social relevance score. Then this paper models the categorical correlations which combine the category bias of users and the popularity of POIs into categorical relevance score. Fourth, this paper effectively match these four relevance scores to generate preference score.Finally, this paper effectively fuse preference score to probabilistic matrix factorization model, generates the recommended list of POIs to user.Experimental results indicate that this model is superior to the advanced NCPD algorithm. On Foursquare dataset, the precision and recall has increased by 27%and 24% respectively. On Twitter dataset, the precision and recall has increased by 26% and 25% respectively. This method significantly improves the accuracy of POI recommendation.2. POI recommendation based on the user check-in behaviorCurrent research lacks an integrated analysis of the joint effect of the geographical influence, temporal effect, social correlation, content information and popularity impact to deal with the problem of data sparsity, especially in the out-of-town recommendation scenario which has been ignored by most existing work. To tackle this challenge, this paper proposes a joint probabilistic generative model to imitate user check-in activities in a process of setting decision, which is the first joint model to integrate the above factors effectively,and exploits geographical correlation to design a good spatial index structure namely spatial pyramid, optimizes smoothly for local preferences, further reduces data sparse problem. This model consists of offline modeling and online recommendation, supports home-town and out-of-town of two recommendation scenario, utilizes an extensible query process technology threshold algorithm to accelerate the online recommendation process.Experimental results indicate that this model is superior to the advanced SVDFeature algorithm. In the out-of-town recommendation scenario, on Foursquare dataset, the precision and recall has increased by 27% and 24%respectively; on Twitter dataset, the precision and recall has increased by 21%and 23% respectively; on Douban dataset, the precision and recall has increased by 22% and 24% respectively. In the home-town recommendation scenario, on Foursquare dataset, the precision and recall has increased by 14%and 16% respectively; on Twitter dataset, the precision and recall has increased by 23% and 20% respectively; on Douban dataset, the precision and recall has increased by 15% and 17% respectively. This method significantly improves the accuracy of POI recommendation.3. POI recommendation based on social media mining and visualizationIn social network based on social media, images haven't been researched very well for POI recommendation. To solve this problem, this paper proposes a social media topic model which jointly models five Twitter features (i.e., text,image, location, timestamp and hashtag), make full use of the intrinsic relationship between these characteristics to build a joint probability generation model. This method researches on the images on Twitter for POI recommendation, solves the problem of noise images, and predefines three criteria: visual relevance, visual coherence, and visual distinctiveness, finally exploits convolutional neural network to select representative images for visualization of POIs. Experimental results indicate that this model is superior to the advanced TRM algorithm. On Twitter dataset, the mean average precision has increased by 22%. This method significantly improves the accuracy of POI recommendation. |
PDF Full Text Request