Recommendations and privacy in the arXiv system: A simulation experiment using historical data

Recommender systems may accelerate knowledge discovery in many fields. However, their users may be competitors guarding their ideas before publication or for other reasons. We describe a simulation experiment to assess user privacy against targeted attacks, modeling recommendations based on co‐access data. The analysis uses an unusually long (14 years) set of anonymized historical data on user‐item accesses. We introduce the notions of “visibility” and “discoverability.” We find, based on historical data, that the majority of the actions of arXiv users would be potentially “visible” under targeted attack. However, “discoverability,” which incorporates the difficulty of actually seeing a “visible” effect, is very much lower for nearly all users. We consider the effect of changes to the settings of the recommender algorithm on the visibility and discoverability of user actions and propose mitigation strategies that reduce both measures of risk.