Contextual-Aware Hybrid Recommender System for Mixed Cold-Start Problems in Privacy Protection

Abstract

With the rapid growth of web services, more and more users elect to hide their information settings. While this practice protects users' privacy, it also creates highly sparse dataset, which in turn causes cold start problems in Recommender Systems (RS) and leads to poor prediction results. Contextual-aware recommender system (CARS) has been recognized as an efficient solution to address either a new user or a new item cold start scenario. It outperforms traditional collaborative filtering recommender systems. Most existing CARS solutions however have not provided efficient solutions for mixed cold start scenarios, in which two or more cold start problems co-exist. In this work, we introduce a Weighted Switching Hybrid Context-Aware (W-SHCA) Recommender System. It is based on two algorithms: Content-based CAMF-CC (Context-Aware Matrix Factorization) and Demographics-based CAMF-CC, and utilizes their weighted sum to perform the prediction when a mixed cold start problem is detected. We exploit the W-SHCA model in three recommendation applications: places of interest (STS), music (Music), and movie (CoMoDa). We illustrate some significant performance differences among these three datasets under various mixed cold start situations, and deduced a stable weight selection pattern for W-SHCA associating with dataset characteristics. We believe that this work is a pioneer in dealing with mixed cold-start problems, and would provide a vital reference point for their future research.