Adversarial Machine Learning: The Case of Recommendation Systems

Abstract

Learning with expert advice framework has drawn much attention in recent years especially in the context of recommendation systems. We consider two challenges that we face in broadly applying this framework in practice. One is the impact of adversarial attack strategies (malicious recommendations) and the other is lack of sufficient recommendation from quality experts (aka sleeping expert setting). In this paper, we discuss some recent results on understanding adversarial strategies and their effect on recommendation systems. In addition, in the sleeping expert setting, we discuss some novel designs for learning alaorithms and the analysis of their convergence properties.