A Sparsity-Aware Hybrid Collaborative Recommendation Method with Multi-Source Data

Abstract

Collaborative filtering (CF) is an effective method, which is widely applied in a great deal of recommendation systems. It can extract latent user preferences and item features from user-item rating matrix, by making use of Matrix Factorization (MF)'s technique to generate recommended items. However, the traditional MF based recommendation methods are faced with sparsity problems caused by unavailable and incomplete user-item rating datasets, which greatly degrades the accuracy of recommendation results. To address the sparsity problem caused by unavailable and incomplete datasets, we propose a sparsity-aware hybrid collaborative recommendation approach with multi-source and heterogeneous data in this paper. Firstly, we adopt more abundant and accessible implicit feedback data rather than explicit user-item rating data to overcome the data unavailability issue. Moreover, to address the “Missing not at random” (MSAR) problem with implicit feedback data, we adopt auxiliary data from user profile or item content to generate more accurate user preference and item features; which will be fused into the MF model to generate accurate recommendation results. More specifically, Principal Component Analysis (PCA) and Stacked Denoising Auto Encoder (SDAE) are used as feature extractor, thereby generating user preferences and item features. Finally, we did a lot of experiments on several datasets to assess the proposed recommendation approach. The results indicate that, compared with pure matrix factorization, the hybrid collaborative recommendation model combining extra latent user/item representation can enhance the recommended results' accuracy; while the time consumption is almost the same.