Hybrid Collaborative Filtering Approaches based on Deep Learning for Recommender System

Abstract

Recommender system has received significant attention from academia and various industries, especially e-commerce service companies in the last decade. Recently, the exploding growth of the number of users and items in e-commerce service companies is causing an extremely high sparseness of relationships between users and items. As a result, the sparsity issue becomes one of the major obstacles to achieving a high performance in collaborative filtering (CF) based recommender system, which mostly relies on the relationships between users and items. To overcome the sparsity issue, several hybrid recommender systems have been proposed, which leverage auxiliary information related to users and items together with the ratings. However, they still fail to effectively exploit auxiliary information, and thus, new approaches are required for recommender systems. In order to effectively exploit auxiliary information, we propose a two-step approach that adopts deep learning methods into CF-based recommender systems. First, to leverage both ratings and documents of items, we develop a novel document context-aware hybrid recommendation model, which integrates Convolutional Neural Network (CNN) into Probabilistic Matrix Factorization (PMF) in order to effectively capture contextual information such as word order or surrounding words of a word in documents of items. Second, to boost the performance of the document context-aware hybrid recommender system, we introduce a new latent factor modeling method for items. The proposed method exploits the statistics of items to make our document context-aware hybrid recommendation model more robust to not only sparse but also skewed datasets. Our experiments will also show that our proposed recommendation models outperform the state-of-the-art recommendation models. The implementations of our models and related datasets will be available at http://dm.postech.ac.kr/~cartopy.