Para 'Normal' Activity: On the Distribution of Average Ratings

In this paper we study the distribution of average user rating of entities in three different domains: restaurants, movies, and products. We find that the distribution is heavily skewed, closely resembling a log-normal in all the cases. In contrast, the distribution of average critic rating is much closer to a normal distribution. We propose user selection bias as the underlying behavioral phenomenon causing this disparity in the two distributions. We show that selection bias can indeed lead to a skew in the distribution of user ratings even when we assume the quality of entities are normally distributed. Finally, we apply these insights to the problem of predicting the overall rating of an entity given its few initial ratings, and obtain a simple method that outperforms strong baselines.