PISA: Combining Transformers and ACT-R for Repeat-Aware Sequential Listening Session Recommendation

Repetitive listening behaviors are often overlooked or inadequately addressed by music streaming platforms which often rely on sequential recommender systems to suggest music based on users’ listening histories. This presents a key challenge as repeatedly listening to the same song over time is a common habit that can shape how users experience and interpret this song. In this paper, we introduce PISA ( P sychology- I nformed S ession embedding using A CT-R), a session-level sequential recommender system designed to overcome this challenge. PISA leverages a Transformer-based architecture to learn embedding representations of listening sessions and users, incorporating attention mechanisms inspired by Anderson’s ACT-R ( A daptive C ontrol of T hought- R ational), a cognitive framework that models human memory and information retrieval. This enables PISA to effectively capture both dynamic and repetitive listening patterns at multiple levels (song and artist level), allowing it to predict which songs users will listen to in subsequent sessions—whether they are familiar repeats or new discoveries. We validate the effectiveness of PISA using publicly available listening data from Last.fm as well as proprietary data from Deezer, a global music streaming platform. Our results highlight the critical role of repetition modeling for sequential listening session recommendation. To support future research in this area, we publicly release our proprietary dataset along with the source code for PISA, facilitating its adoption and further development.