Online news recommender based on stacked auto-encoder

Abstract

Because of the popularity of Internet and mobile Internet, people are facing serious information overloading problems nowadays. Recommendation engine is very useful to help people to reach the Internet news they want through the network. Collaborative filtering (CF), such as item-based CF, is the most popular branch in recommendation domain. But the data's high-dimension as well as data sparsity are always the main problems. A novel CF method is introduced in this article, which uses stacked auto-encoder with denoising, an unsupervised deep learning method, to extract the useful low-dimension features from the original sparse user-item matrices. Together with proper similarity computing algorithms, the method provided in this article is proved to be more precise than the methods based on SVD or item-based CF.