An Adaptive Recommender System for Computational Serendipity

Serendipity is recognized as very challenging to simulate and stimulate in recommender systems. In this paper, we adopt a novel approach to model and implement serendipity in a context of health news recommender system. The proposed conceptual framework for serendipity consists of a surprise component, a value component, and a learning component. The three components work together to reason about what information is serendipitous, defined as both surprising and valuable to a user. The implementation is through a series of computational approaches, resulting a prototype called "StumbleOn". We find that the computational approaches help identifying serendipitous recommendations, which are further improved by adaptively learning users' real-time feedback. This study contributes to the research on how to generate serendipity for users in a predictable and systematic way.