Research On Recommendation Technology In Location-Based Mobile Social Networks

Initially recommender system exploits users’rating information for items, constructs preference model, and takes advantage of content-based recommendation, collaborative filtering, or knowledge-based recommend-dation to generate results for users. Collaborative filtering, which is widely used recommendation algorithm, take the objective phenomenon that two users with similar behaviors probable have similar personal preferences as heuristic rule, and generate recommendation results by synergy between similar users. Furthermore, to solve the data sparsity and cold start problems in collaborative filtering, researchers add users’some social attribute to their preference models, and propose some social recommendation algorithms. However, traditional network platforms are very difficult to provide the full set of social attribute data, such as locations of mobile users.It provides a new way to collect some information including social relations and locations about mobile users by location check-in, information sharing and social activities, also facilitates the development of recommendation system in location-based social networks. In this thesis, some basic approach and key problems about recommendation system in location-based social networks are studied, more details is as follows.(1) Most of traditional user-based collaborative filtering algorithms usually use user-item matrix to compute the similarity for users, while ignore the impact of user’s location. To solve this problem, based on a proposed framework of location-based network services recommendation which combines location-based services with traditional collaborative filtering, propose an approach to compute mobile users’preferences similarity which is based on their geographic location, and prove that it satisfies the general properties of neighbor similarity measure. Then to solve the data sparsity and cold start problems, according with the concept of trust in sociology, a new method is presented for calculating trust value. By importing them into network services recommendation process, an approach of location-based network services recommendation is proposed, which has the characteristics of location-based services and social networks. Make experiments on public dataset MIT, at first we use location-based network services using volatility for all users to analyze the impact of location on users’preferences. Then make an experimental analysis for the impact of users’location and trust relationship on the data sparsity and cold start problem in similarity matrix, finally the proposed algorithm is proved more accurate and feasible in experiments by MAE and P@R.(2) According to users’location mobility, human movements can be divided into two kinds:short-ranged travel and long-distance travel, short-ranged travel is periodic both spatially and temporally and not effected by the social network structure, while long-distance travel seemingly random jumps correlated with social network ties. In view of this, propose a graph model recommendation algorithm based on division of user mobility region. At first, we divide user movement region into local region and remote region, and define local neighbor and remote friend according to overlap between different movement region of all users. Then we construct a user-region-item graph model, assign weight values to all edges between different nodes, and adapt a modification of PageRank algorithm to generate recommendation results. Finally we verify the recommendation accuracy of proposed algorithm and its performance for cold start users using Yelp Academic Dataset.(3) Friend recommendation plays an important role for meeting new friends and enhancing social skills in traditional social networks. It usually recommends potential friends to user by analyzing the structure of social network and mining potential linked edges. Due to emphasis on interest similarity within a certain area, if we still recommend friends to user in location-based social networks, we could not achieve the desired effect. To solve above problem in this thesis, we propose a neighbor friends recommendation algorithm based on check-in location trajectory similarity of mobile users. At first, we extract import location points or regions by analyzing users’check-in data in location-based social networks, define three behavior patterns between two trajectories which are formed in a fixed period and propose a temporal-geographical algorithm of trajectory similarity. Then we can recommend location-based friends to users according to their trajectory similarities. Finally the performances on precision and recall of the proposed algorithm and the positive affect of trust relationship on recommendation results are proved by experimental analysis on real data set from Foursquare.com.(4) In the mobile network, sensing the context changes about mobile users, simulating users’characteristics, automatically forecasting users’ demand and sending right information to the user in the right time and place, are the important research content of mobile recommendation system. In this thesis, to extract mobile users’adaptive options preferen-ces in multidimensional context, we first propose a context-aware learn-ing algorithm for mobile users’adaptive options preferences, which is based on analytical hierarchy process, and the detailed description and complexity of the algorithm are given. Then construct user’s contextual state transition probability matrix according to their state transition in different environment, define and compute contextual state transition similarity among users, and propose two kinds of context-aware mobile proactive recommendation algorithms for normal users and cold-start users. Make experiments on synthetic dataset MobileServices, we first identify which types of contextual information could affect users’ personal preferences by users’average usage volatility for each kind of context. Then the impact of different values of parameter and proportion of training dataset for learning results are completely analyzed. Finally the accuracy of the proposed proactive recommendation algorithm is verified.