基于位置社会网络的双重细粒度兴趣点推荐

Dual Fine-Granularity POI Recommendation on Location-Based Social Networks

兴趣点推荐是在基于位置社会网络(location-based social network,LBSN)中流行起来的一种全新形式的推荐.利用LBSN所包含的丰富信息进行个性化推荐能有效增强用户体验和提高用户对LBSN的依赖度.针对无显示用户偏好、兴趣非一致性和数据稀疏性等挑战性问题,研究一种针对LBSN的双重细粒度POI推荐策略,即一方面将用户的全部历史签到信息以小时为单位细分为24个时间段,另一方面将每个POI细分为多个潜在主题及其分布,同时利用用户的历史签到信息和评论信息挖掘出用户在不同时间段的主题偏好,以实现POI的Top-N推荐.为实现该推荐思路,首先,根据用户的评论信息,运用LDA模型提取出每个POI的主题分布;然后,对于每个用户,将其签到信息划分到24个时间段中,通过连接相应的POI主题分布映射出用户在不同时间段对每个主题的兴趣偏好.为解决数据稀疏问题,运用高阶奇异值分解算法对用户-主题-时间三阶张量进行分解,获取用户在每个时间段对每个主题更为准确的兴趣评分.在真实数据集上进行了性能测试,结果表明所提出的推荐策略具有较好的推荐效果.

Point of interest recommendation is a new form of popular recommendation in locationbased social network （LBSN）. Utilizing the rich information contained in the LBSN to do personalized recommendation can enhance user experience effectively and enhance users dependence on LBSN. Facing the challenging problems in LBSN, such as no explicit user preferences, nonconsistency of interest, the sparseness of data, and so on, a dual finegranularity POI recommendation strategy is proposed, of which, on the one hand, the historical checkin information of each user is divided into 24 time periods in hours; on the other hand, each POI is divided into a number of potential topics and distribution. Both the information of users checkin and comments are used to mine users topic preference in different time periods for Top-N recommendation of the POIs. In order to achieve the recommendation ideas, first of all, according to the comments information on the visited POIs, we use LDA topic generation model to extract the topic distribution of each POI. Secondly, for each user, we divide each users checkin data into 24 time periods, and connect it with the topic distribution of the corresponding POIs to map user interest preference on each topic in different periods. Finally, in order to solve the issue of data sparse, we use higher order singular value decomposition algorithm to decompose the thirdorder tensor of usertopictime to get more accurate interest score of users on each topic in all time periods. The experiments on a real dataset show that the proposed approach outperforms the stateoftheart POI recommendation methods.