Parametric Empirical Bayes for Predicting Quality in Rating Systems
User-solicited ratings systems in online marketplaces suffer from a cold-start problem: new products have very few ratings, which may capture overly pessimistic or optimistic views of the proper rating of that product. This could lead to platforms promoting new products that are actually low quality, or cause high-quality products to get discouraged and exit the market early. In this paper, we address this cold-start problem through an approach that softens early reviews, interpolating them against a background distribution of product ratings on the platform. We instantiate this approach using a parametric empirical Bayes model, which weighs reviews of a new product against expectations of what that product's quality ought to be based on previous products that existed in the same market, especially when the number of reviews for that new product are low. We apply our method to real-world data drawn from Amazon as well as synthetically generated data. In aggregate, parametric empirical Bayes performs better on predicting future ratings, especially when working with few reviews. However, in these same low-data settings, our method performs worse on individual products that are outliers within the population.
READ FULL TEXT