Behavior reflects preference : Mitigating the user cold-start in recommender systems with user telemetry data

Sammanfattning: Recommender Systems are information filtering systems that aim to predict a user’s preference for an item. A central challenge when building a Recommender System is the user cold-start, the integration of new users into the recommendation process. It can currently not be ultimately solved, but only mitigated based on additional information about the user. This work proposes to utilize technical usage data, telemetry data, for user preference modeling. In the industry use-case of an in-game item recommendation system for a mobile game, telemetric features have been engineered, to capture player’s behavior during the first hours inside the game. The prediction of the first purchase was then modeled as a multi-class classification problem. Across a range of different classification model families, the models trained on telemetric features of the present dataset all significantly outperform the same models trained on demographic features, which in turn outperform naive baselines. The result has implications for industry use-cases where Recommender Systems are being employed, and telemetric features can be aggregated, like mobile applications. It also has implications on future research of cold-start mitigation, as telemetric information could be used to generate recommendations in different problem architectures than classification.