Towards Robust and Trustworthy Recommendation Systems

点击翻译按钮获取中文摘要

Recommender systems (RSs) are an integral part of several industries that incorporate choices and strategic behaviours for enhanced decision-making. They reduce the burden of information overload by predicting items of interest to users. In e-commerce, successful recommendations represent tremendous potential for cross-selling and escalated average order sizes per customer, improved customer satisfaction, and controlled consumption for optimizing resources. RSs are powerful tools for Amazon, Walmart, Netflix, and more. In the era of COVID19, by April 21, 2020, U.S. and Canadian e-commerce orders had already exhibited 129% year over year growth, with an impressive 146 % increase in total online retail orders. Widespread use of user-adaptive recommendation systems in open platforms and information access domains, as e-commerce, provides a strong motivation for unscrupulous users to inject malicious profiles "shilling attacks" into the RSs to manipulate recommendation rankings in order to gain economic advantage. For example, Sony Pictures admitted using fake quotes from non-existent movie critics to promote some newly released films. Amazon found that unscrupulous producers were taking a more deceitful artifice, leading to the absurd recommendation results. eBay continually deals with users who subvert the system by purchasing good ratings from other members to bolster their reputations. It is crucial to maintain the fairness of recommendations by constructing a robust system with stable recommendations in the presence of shilling attacks. The focus of this research program is on developing and validating new models and strategies for solving the underlying problems of defending shillings attacks that threaten the recommendation quality, fairness, and manipulate user decisions. In particular, we aim at providing modeling solutions and strategies for two aspects of the attack-resistance process: A) Shilling attack detection; and B) Robustness against malicious attacks. These two aspects are critical for maintaining and increasing the trustworthiness and fairness of RSs. The proposed research program includes the development of effective models and methodologies for discovering such attacks and improving the robustness of RSs using advances in graph theory, clustering analysis, algebraic multigrid, deep learning, and adversarial machine learning. The volume of data produced in the retail industry is expanding exponentially, creating challenges and opportunities for those diligently analyzing these data towards better robustness to gain a competitive advantage. This research program will involve designing new distributed/parallel network architecture and protocols for handling large-scale data using multi-tier networks. Overall, the proposed research aims to solve the above-mentioned challenges for the ultimate establishment of efficient, scalable, and robust RSs through a solid plan of training HQP in highly applicable and cutting-edge techniques.