Users in music streaming platforms typically consume tracks sequentially in sessions by interacting with personalised playlists. To satisfy users, music platforms usually rely on recommender systems that learn users' preferences over individual tracks and rank the tracks within each playlist according to the learned preferences. However, such rankings often do not fully exploit the sequential nature of the users' consumption, which may result in a lower within-a-session consumption. In this paper, we model the sequential within-a-session preferences of users and propose an optimisation-based sequencing approach that allows for optimally incorporating such preferences into the rankings. To this end, we rely on interaction data of a major music streaming service to identify two most common aspects of the users' sequential preferences: (1) Position-Aware preferences, and (2) Local-Sequential preferences. We propose a sequencing model that can leverage each of these aspects optimally to maximise the expected total consumption from the session. We further perform an extensive offline and off-policy evaluation of our model, and carry out a large scale online randomised control trial with 7M users across 80 countries. Our findings confirm that we can effectively incorporate sequential preferences of users into our sequencer to make users complete more and skip less tracks within their listening sessions.
Recent advances in recommender systems (RS) have shown substantial performance gains through generative modelling. In practice, recommendation often involves constructing slates--ordered lists of items--that must satisfy multiple objectives beyond relevance, such as constraints defined over item attributes or fairness constraints. Existing multiobjective approaches either rely on post-processing techniques designed for non-generative settings, or incorporate auxiliary objectives directly into model training. The former does not explicitly account for the sequential nature of generative RS, while the latter is often impractical in large-scale systems. We propose a lightweight, inference-time decoding layer that augments autoregressive generative RS to support multiobjective slate generation without modifying or retraining the underlying model. We formulate decoding as an online constrained optimisation problem, where items are selected sequentially, and trade-offs between relevance and auxiliary objectives are adjusted dynamically based on the remaining constraint slack, i.e., how much of each objective remains to be satisfied. This is implemented via a stochastic primal-dual approximation scheme that balances relevance and auxiliary objectives during generation. We provide theoretical guarantees on constraint violation and regret, and evaluate the proposed approach through extensive offline experiments and a large-scale online A/B experiment in a real-world recommender system. Our results show consistent improvements in multiobjective trade-offs, including a +1.8% gain in the auxiliary objectives achieved at zero cost to user satisfaction.