Visualizing ratings in recommender system datasets

The numerical outcome of an offline experiment involving different recommender systems should be interpreted also considering the main characteristics of the available rating datasets. However, existing metrics usually exploited for comparing such datasets like sparsity and entropy are not enough informative for reliably understanding all their peculiarities. In this paper, we propose a qualitative approach for visualizing different collections of user ratings in an intuitive and comprehensible way, independently from a specific recommendation algorithm. Thanks to graphical summaries of the training data, it is possible to better understand the behaviour of different recommender systems exploiting a given dataset. Furthermore, we introduce RS-viz, a Web-based tool that implements the described method and that can easily create an interactive 3D scatter plot starting from any collection of user ratings. We compared the results obtained during an offline evaluation campaign with the corresponding visualizations generated from the HetRec LastFM dataset for validating the effectiveness of the proposed approach.