Protecting the Privacy of Observable Behavior in Distributed Recommender Systems

摘要: Implicit and explicit evidence can be thought of in a unified framework as “observable behavior;” rating is merely one type behavior [1]. Adequate protection privacy prerequisite to the use observable (a necessary, but not sufficient condition—a willingness help shape information space also required, for example). Privacy protected cases where intended public (e.g., building Web links) or when users believe that adequate safeguards exist Amazon’s purchase make recommendations). We are interested exploring distributed techniques used centralized services unable establish acceptable safeguards. In remainder this paper we U = user, I item, B behavior, R recommendation, F feature. Centralized recommender systems based on implicit feedback often map from U×I×B array observations U×I matrix recommendations. This then form either an I×I item similarity basis “cross-selling,” U×U user find with similar tastes, which recommendations constructed using matrix. systems, sharing about potentially problematic, even pseudonyms used, because side could serve pinpoint identity individual (e.g, there may only pilot who lives College Park does retrieval research) [2]. Some ways suffer weakness, models preferences scalar relationships, while true preference relationships individualized aggregate multiple factors. therefore define more abstract I×F shared making At point, no commitment meaning features. require able compute personally useful I×R matrix, they update their personal (and private) I×B A restricted case would treat items independent mappings R. For example, (read 15 seconds, forwarded boss, saved) might (related retrieval, high quality), turn (high interest at work, low home). Each community must agree common , nature need completely standardized.