Recommendation: A Less Explored Killer-App of Uncertainty?

Due to the unprecedented amount of information available, it is becoming more and more important to provide personalized recommendations on data, based on past user feedbacks. However, available user feedbacks or ratings are extremely sparse, which motivates the needs for rating prediction. The most widely adopted solution has been collaborative filtering, which (1) identifies "neighboring" users with similar tastes and (2) aggregates their ratings to predict the ratings of the given user. However, while each of such aggregation involves varying levels of uncertainty, e.g., depending on the distribution of ratings aggregated, which has not been systematically considered in recommendation, though recent study suggests such consideration can boost prediction accuracy. To consider uncertainty in rating prediction, this paper reformulates the collaborative filtering problem as aggregating community ratings into multiple predicted ratings with varying levels of certainty, based on which we identify top-k results with both high confidence and rating. We empirically study the efficiency and accuracy of our proposed framework, over a classical collaborative filtering system.