Recommender systems based on quantitative implicit customer feedback

Due to the abundant variety of products offered by e-commerce companies and online service providers, recommender systems become an increasingly important decision aid for customers. In this paper we focus on quantitative implicit customer feedback like sales and play records data. We extend the current state-of-the-art method for recommendations based on matrix factorization under a normal distribution assumption by allowing for different distributions which are more suitable to model this kind of data. In particular, we use the Poisson, the inverse Gaussian and the gamma distribution as extensions. The experimental evaluation with three real-world data sets shows the improved performance of our approach and we demonstrate the merit of using various distributions depending on the respective data set.