Trust-aware Collaborative Denoising Auto-Encoder for Top-N Recommendation

Both feedback of ratings and trust relationships can be used to reveal user preference to improve recommendation performance, especially for cold users. However, the high-order correlations between tow kind of data are always ignored by existing works. Towards this problem, we propose a Correlative Denoising Autoencoder (CoDAE) model to learn correlations from both rating and trust data for Top-N recommendation. First, a novel deep learning model CoDAE, in which two mid-layers from separate stack denoising autoencoders are fused into one shared layer. Advancing previous works which utilize these data in shallow level, this model can effectively extract high-order correlations from low-level representations of these data for recommendation. Second, to further learn implicit corrections between these two autoencoders, we develop a novel correlative regulation to build the relations between other hidden layers of the two separate autoencoders. In this way, this model can learn correlations more effectively and thus improve the recommendation quality. Comprehensive experiments on two public datasets demonstrate that the CoDAE significantly outperforms other state-of-the-art approaches in top-N recommendation task.