The Design and Analysis of Location-Based Collaborative Filtering Policies for Mobile Recommender Systems

• Main Authors: Chi-Fu Hung, 洪啓富
• Other Authors: Chiu-Ching Tuan
• Format: Others
• Language: en_US
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/q68y6a

博士 === 國立臺北科技大學 === 電子工程系博士班 === 105 === 　　The popularity of smartphones and mobile applications has increased, and location-based service (LBS) applications in particular have been widely used recently. Traditional LBS recommender systems for mobile users (MUs) apply a complete rating database to support LBS environments and approach the users’ preferences toward points of interest (POIs), after evaluating the ranks of interest toward the recommended POIs. However, such recommender systems seem to ignore the users’ moving direction and speed, and they may thus recommend POIs that are behind the current positions of the MUs; therefore, the users are often unwilling to return to the positions that they have already traversed. Determining how to efficiently and accurately calculate POIs by considering the moving directions and patterns of MUs has become an essential research topic in recent years. 　　This dissertation proposes three recommender systems, namely the hybrid rating location-based collaborative filtering system (HRLCF), the location-based social networks collaborative filtering system (LBSNCF), and the location-based collaborative filtering with dynamic time periods system (LCFDTP). The HRLCF applies a position prediction algorithm to determine location-based recommendations by considering the moving directions and speeds of MUs. The system actively analyzes the usage behavior patterns of MUs and then evaluates their preferred weighting for every recommended item. Finally, it recommends the top-N POIs being with high weightings to the users. The LBSNCF also applies the position prediction algorithm as used in the HRLCF to predict users’ movements. In particular, the LBSNCF predicts the ratings of POIs that have not yet been rated by evaluating the explicit and implicit ratings of POIs provided by the users’ social network friends. Hence, to calculate the POI weightings, the system analyzes the POI ratings and the relative position and distance between the POIs and the users. Finally, it recommends the top-N POIs to the users. The LCFDTP establishes a query range in which the GPS coordinates of users are considered to be the center; the system then identifies POIs in the neighborhood that have been recently rated by all users. A POI recency filtering algorithm is applied to screen POIs that do not meet users’ temporal conditions and to reduce the time required for calculating the similarity of user ratings. Expected factors of the relative distances between users and POIs in different directions are computed. The system thus can promptly obtain top-N items that closely match the users’ current space–time conditions. 　　This study simulated the three location-based collaborative filtering (CF) systems to evaluate their performance. The proposed systems were compared with three existed recommender systems according to the three evaluation metrics: recommendation accuracy, recommendation coverage, and average recommendation time. The experimental results showed that the HRLCF yielded higher levels of recommendation accuracy and coverage than the other three systems did. However, the average recommendation time of the HRLCF was slightly increased, but this can be acceptable for users. The LBSNCF and LCFDTP achieved higher levels of recommendation accuracy and coverage and a shorter average recommendation time than the other three systems did. 　　The proposed HRLCF, LBSNCF, and LCFDTP were compared, and the results revealed that the recommendation accuracy of the LCFDTP was, on average, 39%–54% higher than those of the HRLCF and LBSNCF. The main reason is that the LCFDTP filters POIs that have not been recently rated and that the LBSNCF applies a user’s friends’ ratings for the recommendation process, thus improving their recommendation accuracy. In addition, the recommendation coverage of the HRLCF was approximately 2%–13% higher than those of the LCFDTP and LBSNCF. The main reason is that the LBSNCF applies a user’s friends’ ratings, and the similarity of preferences of the user and his or her friends are very high. Moreover, the LCFDTP does not recommend unpopular POIs because it filters out POIs that have not been recently rated. This simplifies the calculation of similarity of users’ POI ratings, thus reducing the average recommendation time the LCFDTP required. The average recommendation time of the LCFDTP was approximately 25%–71% shorter than those of the HRLCF and LBSNCF. LBSNCF simplifies the calculations of similarity by using friends’ ratings, but numerous unrated items still exist, thus rating similarity must be calculated through traditional calculation approaches. The excessively long average recommendation time of the HRLCF is due to the numerous precision calculations and the complexity of similarity calculation. Therefore, selecting an appropriate number of POIs and simplifying the similarity calculation for unrated items will be the key factors for achieving the most favorable recommendation accuracy, recommendation coverage, and average recommendation time.