Learning pairwise comparisons of items with bigram content features for recommending

In general, users usually rate items according to interestingness of some features in items on the internet. Considering competitive relationships of ratings on one user interest level and context information of the item content features, this paper proposes an approach to predict items' ratings basing on paired comparisons of different rating items with bigram content features. In the paper, we assume that the user interest on each item can be represented by the combination of different bigram content features, and employ Bradley-Terry model to confirm the user interestingness of each feature pair. Experimental results show that this approach outperforms popular approaches and the competitive approach without context information.