Exploiting Ranking Consistency Principle In Representation Learning For Location Promotion
DATABASE SYSTEMS FOR ADVANCED APPLICATIONS (DASFAA 2018), PT II(2018)
摘要
Location-based services, which use information of people's geographical position as service context, are becoming part of our daily life. Given the large volume of heterogeneous data generated by location-based services, one important problem is to estimate the visiting probability of users who haven't visited a target Point of Interest (POI) yet, and return the target user list based on their visiting probabilities. This problem is called the location promotion problem. The location promotion problem has not been well studied due to the following difficulties: (1) the cold start POI problem: a target POI for promotion can be a new POI with no check-in records; and (2) heterogeneous information integration. Existing methods mainly focus on developing a general mobility model for all users' check-ins, but ignore the ranking utility from the perspective of POIs and the interaction between geographical and preference influence of POIs.In order to overcome the limitations of existing studies, we propose a unified representation learning framework called hybrid ranking and embedding. The core idea of our method is to exploit the ranking consistency principle into the representation learning of POIs. Our method not only enables the interaction between the geographical and preference influence for both users and POIs under a ranking scheme, but also integrates heterogeneous semantic information of POIs to learn a unified preference representation. Extensive experiments show that our method can return a ranked user list with better ranking utility than the state-of-the-art methods for both existing POIs and new POIs. Moreover, the performance of our method with respect to different POI categories is consistent with the hierarchy of needs in human life.
更多查看译文
关键词
Location promotion, Ranking consistency, Graph embedding
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络