MLE: A general multi-layer ensemble framework for group recommendation

Abstract

As the number of users and locations has increased dramatically in location-based social networks, it becomes a big challenge to recommend point-of-interests (POIs) meeting users’ preference. In traditional recommendation tasks, personalized recommendations performs well, however, these methods also have many disadvantages such as the long-tailed problem and the strong assumption. Further, in general scenarios, a group of users (e.g., colleagues, friends, and family members) often visit a specific location to enjoy time together (e.g., meal and shopping). Thus, it is more meaningful to recommend locations to the group than to individuals. However, the existing group recommendation approaches also have some limitations that hardly capture the preferences of a group of users effectively. To make full use of the users’ preferences and improve the effectiveness of group recommendation, in this paper, we propose a multi-layer ensemble framework which has a two-step fusion process. For the first step, we employ several personalized recommendation methods to generate the recommendations for individuals, and the recommendation list is obtained using the proposed fusion approach based on the supervised learning. For the second step, we utilize several ranking aggregation algorithms to fuse the recommendations list of individuals in the group and propose an unsupervised learning based ranking algorithm (URank) to further fuse the results of ranking aggregations to obtain the final group recommendation list. The experiments are conducted on a real-world dataset, and the results demonstrate the effectiveness of our proposed general framework.