Two-phase layered learning recommendation via category structure

Context and social network information have been introduced to improve recommendation systems. However, most existing work still models users' rating for every item directly. This approach has two disadvantages: high cost for handling large amount of items and unable to handle the dynamic update of items. Generally, items are classified into many categories. Items in the same category have similar/relevant content, and hence may attract users of the same interest. These characteristics determine that we can utilize the item's content similarity to overcome the difficultiess of large amount and dynamic update of items. In this paper, aiming at fusing the category structure, we propose a novel two-phase layered learning recommendation framework, which is matrix factorization approach and can be seen as a greedy layer-wise training: first learn user's average rating to every category, and then, based on this, learn more accurate estimates of user's rating for individual item with content and social relation ensembled. Based on two kinds of classifications, we design two layered gradient algorithms in our framework. Systematic experiments on real data demonstrate that our algorithms outperform other state-of-the-art methods, especially for recommending new items.