Mitigating Power User Attacks on a User-Based Collaborative Recommender System.
David C. Wilson, Carlos E. Seminario · The Florida AI Research Society · 2015
Collaborative Filtering (CF) Recommender Systems (RSs) ease the burden of information overload faced by online users who browse, search, or shop for products and services. Influential users, known as “power users” are able to exert substantial influence over recommendations made to other users, and RS operators encourage the existence of power user communities to help fellow users make informed purchase decisions. However, the influence power users wield can be used for both positive (addressing the “new item” problem) or negative (attack) purposes. Attacks on RSs tend to bias recommendations by introducing fake reviews or ratings and remain a key problem area for system operators. In prior work, we have shown that attackers emulating power users are effective against user-based, item-based, and SVD-based CF RSs. Previous research has shown that, in general, attacks on RSs can be mitigated by detecting the attackers and either removing them from the dataset or ignoring them during the prediction calculations. In live RS environments, however, these approaches impact legitimate users detected as attackers (false positives) and can lead to reduced coverage for those legitimate power users as well as reduced accuracy for other users that depend on the ratings of those legitimate power users. Our research is investigating alternative mitigation approaches to address these issues for power user attacks. We focus on techniques that remove or reduce the influence of power users and determine their impact on RS accuracy and robustness using established metrics. We introduce a new metric used to assess the trade-off between accuracy and robustness when our mitigation approaches are applied. And our results show that, for user-based systems, reducing power user influence is more effective than removing power users from the dataset.