Online recommender systems are widely deployed as tools to guide users towards items they will like. There is a growing concern that recommender systems may be manipulated by people with a vested interest in having certain items recommended (or not recommended). This is exacerbated as it is often easy for a manipulator to create multiple online accounts to execute an attack. The goal of this project is to develop general techniques for the design of manipulation-resistant recommender systems as well as specific solutions for applications in which such a recommender could have a significant impact. The applications studied include Internet sites that use user-provided ratings or tags to recommend items to users, and a recommendation system for job candidates that aggregates informal information from former employees and colleagues.
The project uses three research methods: (1) Theoretical modeling and analysis of provably robust mechanisms, building on techniques from economic mechanism design and online learning; (2) Simulations of such mechanisms against known attack strategies; and (3) Empirical tests on recommendation datasets. The results and software will also be used to enrich a Masters-level course on recommender systems, and related courses.
A manipulation-resistant recommender algorithm will be valuable to numerous Internet organizations, allowing them to offer reliable recommendations without obtaining and verifying personal information. The development of a recommendation mechanism for job candidates will directly benefit both employers and candidates, especially those who lack formal credentials. Algorithms and simulation software developed through this project will be made publicly available via the project Web site (http://icd.si.umich.edu/MRRS) that will include additional information on the project.
