Unifying recommendation and active learning for information filtering and recommender systems

The enormous scale of the available information and products on the Internet has necessitated the development of algorithms that intermediate between options and human users. These AI/machine learning algorithms attempt to provide the user with relevant information. In doing so, the algorithms may incur potential negative consequences stemming from the need to select items about which it is uncertain to increase predictive accuracy versus the need to select items about which it is certain to increase recommendation accuracy. This tension between predicting relevant recommendations to the users and learning about the user's interests can be considered an instantiation of the well-known exploration-exploitation tradeoff in the context of information filtering and recommender systems. Building from existing machine learning algorithms, we introduce a parameterized model that unifies and interpolates between recommending relevant information and active learning. We present three experiments investigating the unified model. Specifically, we illustrate the tradeoffs of optimizing prediction and recommendation within a tightly controlled concept-learning paradigm, show the conditions under which a broad parameter range can optimize for both, and identify the effects of human variability on algorithm performance. Thus, combining methods and models from cognitive science and computer science, we quantify implications of tradeoffs between recommendation accuracy and learning about preferences of human users, demonstrating the value of experimental approaches to understanding real world human-machine feedback loops.