Detecting Vicious Users in Recommendation Systems

摘要

Spam and noisy ratings affect the performance of recommendation systems which can lead to incorrect estimations and predictions. The challenge is to discover noisy ratings early in order to isolate its impact. In this paper we suggest an analysis using positive feedback which considers the user's level of confidence, and grades the user from completely honest to complete dishonest. The calculated user's level of confidence is computed based upon the detected level of honesty and affect his ratings. Each domain of ontologies has a calculated region of rejection and non-rejection using each user confidence level, placing his ratings in one region or another and thereby affecting his level of confidence. We used a Movie Lens of 1M ratings dataset to perform the required training. Suggested method has distinguished perfectly between Normal, Excess, Inferiority, and completely dishonest.