Learning to context-aware recommend with hierarchical factorization machines

In recent years, a considerable number of context-aware recommendation methods have been proposed. One such technique, the Factorization Machines (FM) model, estimates interactions among features by working with any real valued feature vector. Although the FM model can efficiently handle arbitrary relationships (i.e., dyad, triad, etc.), it still has its limitations. In real world scenarios, contextual features can be considered as being organized in an understandable and intuitive hierarchy. However, existing the FM model performs poorly with regard to exploiting the hierarchical properties of contextual features during prediction. In this study, we consider the problem of exploiting hierarchical structures to improve recommendation quality and propose a novel two-stage recommendation model called Hierarchical Factorization Machines (HFM). In the first stage of HFM, the proposed model estimates the FM model parameters locally for each tree node and returns the initial predictions at all resolutions. Then, it finely tunes these predictions globally through a tree-structured Markov model. In the second stage, model fitting is achieved through an Expectation-Maximization (EM) algorithm, wherein the generalized Kalman filtering algorithm is used in the inner loop. Extensive experiments on real datasets verify that the proposed model is efficient and effective.