Multi-view factorization machines for mobile app recommendation based on hierarchical attention

Abstract

Mobile app recommendation has been an effective solution to overcoming the information overload in mobile app markets. Recent studies have demonstrated the power of neural network in recommendation tasks which is however rarely exploited for mobile apps. As one of the development of neural network, attention-based models have shown promising results for recommendation because of its capability of filtering out uninformative features from raw inputs. In this paper, to effectively predict users’ preferences for apps, we propose a hierarchical neural network model called MV-AFM for app recommendation which models the interactions of features from different views (view interactions for short) through the attention mechanism. Specifically, the novelty of MV-AFM is the introduction of view segmentation for feature interactions and the construction of two level attention networks: the feature-level attention, starting from the feature embeddings within each view, which intends to select the representative features for the view, and the view-level attention, which learns the importance of interactions between any two views. Extensive experiments on two real-world mobile app datasets demonstrate the effectiveness of MV-AFM.