Optimal recommendation sets: covering uncertainty over user preferences

摘要: We propose an approach to recommendation systems that optimizes over possible sets of recommended alternatives in a decision-theoretic manner. Our selects the alternative set maximizes expected valuation user's choice from set. The set-based optimization explicitly recognizes opportunity for passing residual uncertainty about preferences back user resolve. Implicitly, chooses with diversity optimally covers preferences. can be used several preference representations, including utility theory, qualitative models, and informal scoring. develop specific formulation multi-attribute which we call maximization max (MEM). go on show this is NP-complete (when are described by discrete distributions) suggest two efficient methods approximating it. These approximations have complexity same order as traditional k-max operator and, both synthetic real-world data, perform better than recommending k-individually best (which not surprise) very close optimum less expected).