Stacked Denoising Autoencoder-based Deep Collaborative Filtering Using the Change of Similarity

Recommender systems based on deep learning technology pay huge attention recently. In this paper, we propose a collaborative filtering based recommendation algorithm that utilizes the difference of similarities among users derived from different layers in stacked denoising autoencoders. Since different layers in a stacked autoencoder represent the relationships among items with rating at different levels of abstraction, we can expect to make recommendations more novel, various and serendipitous, compared with a normal collaborative filtering using single similarity. The results of experiments using MovieLens dataset show that the proposed recommendation algorithm can improve the diversity of recommendation lists without great loss of accuracy.