Collaborative filtering for mobile application recommendation with implicit feedback

Abstract

This paper introduces a novel dataset regarding the installation of mobile applications in users devices, and benchmarks multiple well-established collaborative filtering techniques, leveraging on the user implicit feedback extracted from the data. Our experiments use 3 snapshots provided by Aptoide, one of the leading mobile application stores. These snapshots provide information about the installed applications for more than 4 million users in total. Such data allow us to infer the users activity over time, which corresponds to an implicit measure of interest in a certain application, as we consider that installs reflect a positive user opinion on an app, and, inversely, uninstalls reflect a negative user opinion. Since recommendation systems usually use explicit rating data, we have filtered and transformed the existing data into binary ratings. We have trained several recommendation models, using the Surprise Python scikit, comparing baseline algorithms to neighborhood-based and matrix factorization methods. Our evaluation shows that SVD-based and KNN-based methods achieve good performance scores while being computationally efficient, suggesting that they are suitable for recommendation in this novel dataset.