Detecting sparse rating spammer for accurate ranking of online recommendation

Abstract

Ranking method for online recommendation system is challenging due to the rating sparsity and the spam rating attacks. The former can cause the well-known cold start problem while the latter complicates the recommendation task by detecting these unreasonable or biased ratings. In this paper, we treat the spam ratings as 'corruptions' which spatially distribute in a sparse pattern and model them with a L1 norm and a L2,1 norm. We show that these models can characterise the property of the original ratings by removing spam ratings and help to resolve the cold start problem. Furthermore, we propose a group-reputation-based method to re-weight the rating matrix and an iterative programming-based technique for optimising the ranking for online recommendation. We show that our optimisation methods outperform other recommendation approaches. Experimental results on four famous datasets reveal the superior performances of our methods.