Abstract de la publi numéro 19536

Song recommendation from listening counts is now a classical problem, addressed by different kinds of collaborative filtering (CF) techniques. Among them, Poisson matrix factorization (PMF) has raised a lot of interest, since it seems well-suited to the implicit data provided by listening counts. Additionally, it has proven to achieve state-of-the-art performance while being scalable to big data. Yet, CF suffers from a critical issue, usually called cold-start problem: the system cannot recommend new songs, i.e., songs which have never been listened to. To alleviate this, one should complement the listening counts with another modality. This paper proposes a multi-modal extension of PMF applied to listening counts and tag labels extracted from the Million Song Dataset. In our model, every song is represented by the same activation pattern in each modality but with possibly different scales. As such, the method is not prone to the cold-start problem, i.e., it can learn from a single modality when the other one is not informative. Our model is symmetric (it equally uses both modalities) and we evaluate it on two tasks: new songs recommendation and tag labeling.