Efficient Promotion Algorithm by Exploring Group Preference in Recommendation

Abstract

Recommendation is an important issue in e-commerce systems. Conventional recommender algorithms, like the collaborative filtering recommendation algorithms, have been extensively studied and developed into a very mature stage, in which how to further promote user's favorite items and alleviate sparsity and cold start problem become increasingly important. In this paper, traditional recommender algorithms are promoted by exploring group's preference to mitigate the issues above and improve the predicting accuracy in both cases. Our work is based on the following observation. Users in the same group share common interests, given a group there are items that the group is most interested in, on the other hand, given some specific items there is the first-rate group which shows most preference compared to other groups. This leads to our proposed PromoRec algorithm which focuses on promoting items that users are most likely to prefer with sparse insensitivity since group enriches user's data largely. In a multi-dimensional space, we show how to efficiently compute (a) the most popular items for a target group and (b) the group which shows most interests in specific items. In addition, without the needs of available user group information, we propose an automatic classification algorithm based on users' similar interests. To improve the recommendation accuracy, we use additional item classification information to help determine the similarity between users. The experiment results confirm that our method significantly enhanced the traditional item recommendation algorithms especially while predicting ratings of promoted items for sparse users.