A robust recommended system based on attack detection

Abstract

SummaryIn the increasingly fierce competition in e‐commerce sites, the recommendation system has brought great benefits to the site, but some unscrupulous businesses use the recommended system algorithm loopholes, the use of bulk injection of some fake users, and the ratings of these users with the normal user's rating. Therefore, when calculating user similarity, it is easy to enter the user's neighborhood circle, because the false user takes a high score (“push attack”) or a low score (“null attack”) on the target project. The recommendation scores will be biased, and it is important to detect these false users for the recommendation system and maintain a good e‐commerce competitive environment. In this paper, we propose a recommendation algorithm that divides user‐generated ratings into normal ratings and non‐generic ratings and refine and use state information to minimize the impact of spurious malicious users. Our algorithm first ensures that the recommender system is stable against the three main attack modes (random attack, average attack, power flow attack). Through the analysis of the real data, we verify the performance of the proposed scheme and compare our algorithm with the existing one.