Context-aware recommendations via a tree-based ensemble framework

Incorporating contextual information is very useful to improve the accuracy of personalized recommendations. However, how to utilize the contexts more efficiently are always confusing researchers, the existing work ignores the connections among different contexts, and even suffers from data sparsity. Cross-context behavior transfer has been tested to help overcome the problems. In this paper, we propose a tree-based ensemble framework of exploiting behavioral relations between different contexts to make context-aware recommendations. Specifically, our framework is to make recommendations for the target context based an ensemble of different user/item feature vectors learnt from other contexts with a regularization term. The empirical result and analysis on real-life data demonstrate that our framework achieves a significant increase in recommendation accuracy.