Opredelitev in predikcija uporabnikovega načina poslušanja na pretočnih platformah iz metapodatkov o interakciji

As streaming services have made music ubiquitous, users increasingly expect their listening experiences to adapt to their current situation. While contextual recommendation systems aim to meet this need, they typically rely on external context and overlook internal factors such as emotional state. Consequently, although generally seen as valuable, they often fail to align with user expectations. Motivated by user-study-centered research, which shows that people engage with music through distinct interaction paradigms, we propose that listening intent can be inferred directly from behavior. Using a Spotify dataset, we cluster listening sessions based on interpretable interaction features to identify listening modes and predict them early in a session. Our analysis reveals three distinct modes. Passive listening involves minimal interaction. In active exploration, users navigate around the platform, manually exploring tracks, while active refinement sees targeted skipping to fine-tune existing collections. Passive listening may be related to music discovery, whereas active refinement may reflect live curation for a particular mood, though additional data is needed to confirm this. We show that listening modes can be predicted with promising accuracy, improving as more interactions are observed. Contrary to some research, we find that while background listening is frequent, active refinement is even more prevalent, and that skipping may sometimes reflect curation rather than dissatisfaction. These findings can help streaming platforms anticipate user expectations and adapt their interfaces accordingly. Beyond system design, our work provides data-driven insights relevant to user-centered fields, such as human-computer interaction and multimedia, and illuminates how users engage with music in the streaming era.