Mining User Preferences for Recommendation: A Competition Perspective

Mining user preferences plays an important role in building personalized recommender systems. Instead of mining user preferences with the item content or the user-item-rating matrix, we exploit Bradley-Terry model to mine user preferences as pairwise comparisons. In this paper we assume that the user preference on each item can be represented by the combination of different content features, which brings a direct bridge between features and user preferences. Experimental results show that the method based on pairwise comparisons outperforms baseline approaches with less recommendation time.