Research Internship - Music Recommendation (m/f/d)

Research Topic -- Better than average: how to represent artists embeddings from their tracks embeddings?

Most modern recommender systems rely on a compressed representation to model items: embeddings. A profusion of techniques may provide such representations: matrix factorisation (e.g., diagonalisation, SVD, NMF), word2vec-like sequence embedders, or as a by-product of a deep neural network learning task (e.g., VAE, Transformer). Said embeddings are often latent, and their coordinates do not have a direct interpretation and meaning; namely, it is only taken as a whole that the structure and relative positions of embeddings start to make sense.  The goal of this internship is to explore embedding aggregation strategies beyond the widespread choice of the average, or, alternatively, to structure the embedding space in ways that make the average make sense. The underlying application is to be able to represent several types of items (tracks, albums, artists, playlists, user history…) in a unified way

This internship has a start date on February onwards and has a duration of 6 months.