Modeling Musical Genre Trajectories through Pathlet Learning
This paper investigates the dynamic evolution of users’ music preferences over time. To model such temporal patterns, we propose *pathlets*—interpretable and reusable units representing characteristic listening trajectory motifs. We introduce dictionary learning to music trajectory modeling for the first time, automatically discovering cross-user common evolutionary pathways from genre sequences. By jointly optimizing a pathlet dictionary and trajectory embeddings, our approach yields compact, semantically meaningful, and generalizable temporal preference representations. Evaluated on a real-world dataset from Deezer comprising 2,000 users and 17 months of genre-level behavioral data, the learned pathlets effectively uncover canonical listening evolution patterns—including genre expansion, reversion, and migration—substantially enhancing trajectory interpretability and improving downstream diversity-aware recommendation performance. Our framework establishes a novel paradigm for dynamic music preference modeling and personalized service delivery.