Accepted Papers

Author: Dhyey Desai

Abstract: Music recommendation systems shape which artists listeners encounter and, by extension, which artists receive attention in highly concentrated streaming environments. Prior work has shown that music recommenders can reproduce and amplify gender imbalance in artist exposure. This paper studies artist gender bias in a collaborative-filtering pipeline built on the Last.fm 360K dataset. Using 24,509 users, 10,362 artists, and 3,867 gender-labeled artists sourced from MusicBrainz, we train an Alternating Least Squares (ALS) recommender and evaluate it with both standard ranking metrics and exposure-oriented fairness metrics. We then compare three mitigation strategies: (1) post-processing score boosting, (2) a novel fairness-regularized ALS (Fair-ALS) that corrects bias during training via iterative confidence reweighing, and (3) a stacked approach combining both. The baseline model allocates 69.8% of position-weighted exposure to male artists. Post-processing alone (Mitigation B, +28% boost) reduces this to 49.2% but at a 1.6% NDCG cost. Fair-ALS (𝜆 = 1.0) reduces male exposure to 62.3% with zero NDCG cost. The stacked approach achieves true gender parity (49.97% male share) while reducing NDCG cost by 30% relative to post-processing alone, demonstrating a Pareto improvement across the full fairness–utility spectrum. Bootstrap 95% confidence intervals confirm all differences are statistically distinguishable.

Authors: Timothy Greer, Wo Jae Lee, Aileen Mi, Emanuele Coviello and Amina Shabbeer

Abstract: Music recommendations rely on robust artist embeddings. We introduce MusicContainerNet (MCN), a lightweight self-attention model that aggregates an artist's track embeddings---capped at 20 tracks---into a single artist representation. The cap keeps inference fast (~5,000 artists/second on one GPU) and meets Amazon Music's strict latency budget. Offline, MCN variants substantially outperform averaging baselines on retrieval and tagging tasks. In a live A/B test, MCN drove +4 bps in Play Days, +8 bps in Listening Hours, and +11 bps in Long-Term Engagement Retention relative to the production baseline, validating real-world user impact.

Authors: Marta Moscati, Emra Gurung, Shah Nawaz, Monorama Swain and Markus Schedl

Abstract: Emotions constitute an important aspect when listening to music. This has motivated the development of high-quality emotion characterizations of music tracks that can be used for music recommender systems (MRS) that rely on emotion information. Although previous studies evaluate the accuracy of emotion-informed MRS, surprisingly there are no studies that analyze whether music recommendations reflect the distribution of emotions over music catalogs and, more importantly, over users' past listening history. In this work we show that MRS tend to over-represent some emotions in the recommendation lists, compared with both the music catalog and the users' listening history. With this insight, we propose the use of emotion calibration in MRS and show that this has no substantial impact on recommendation accuracy. Code: https://github.com/mmosc/emotion_calibration

Authors: Zhonghao Luo, Jianhong Chen, Nathan Apolonio and Amina Shabbeer

Abstract: We present a production-scale playlist curation system on Amazon Music that leverages modular architecture and holistic LLM-powered reasoning to generate playlists. Our system leverages LLMs to reason over rich track metadata---including genre, mood, era, sonic descriptions, and artist context---to select cohesive track sets that satisfy both relevance and coherence criteria. We formulate the curation task under two complementary inference paradigms: track-wise (binary classification) and list-wise (subset selection). We systematically compare these paradigms and evaluate eight prompt variants across three design dimensions: playlist-level reasoning, track-level reasoning and output structure. Our prompt optimization establishes that outputting only negative tracks achieves 0.96 F1 with 84% fewer output tokens compared to naive all-track approaches. To meet production latency and cost constraints, we develop a teacher-distillation pipeline that transfers reasoning capabilities from Claude Opus 4.6 to Qwen3.5-4B via supervised fine-tuning, achieving comparable playlist quality with 227x cost reduction. Evaluation via recreating both expert-curated editorial playlists and real user personalized playlists demonstrates consistent improvements over production baselines across relevance metrics (weighted precision, weighted NDCG) and coherence metrics (consumption similarity, valence deviation, danceability deviation), confirming that LLMs can serve as effective quality gates for algorithmic playlist curation at scale.

Authors: Jianhong Chen, Aileen Mi, Amina Shabbeer, Zhonghao Luo, Tim Greer and Ajita Mishra

Abstract: Personalized playlists are a core experience in music streaming, yet generating titles and descriptions that are both representative of playlist content and evocative beyond generic templates remains an unsolved challenge at scale. We show that high-quality explanations can be generated largely offline through precomputed reasoning, eliminating expensive online LLM inference without sacrificing quality. We begin by evaluating Qwen3.5-4B, a compact model that substantially reduces per-request cost but retains two fundamental limitations: inference is repeated independently for each playlist without amortizing over reusable content, and the model’s limited capacity imposes a quality ceiling. To progressively address these shortcomings, we propose two methods: Nearest Seed Lookup (NSL), which eliminates online inference entirely by retrieving explanations from a precomputed nearest-seed index, and Two-Stage Atom Composition (2SAC), which decouples reasoning from the serving path by having a large model (Qwen3.5-35B) generate per-track descriptive “atoms” offline while a small model (Qwen3.5-0.8B) assembles them into coherent explanations at serving time. In a seven-criteria LLM-as-a-Judge evaluation over 1,000 playlists, NSL eliminates online inference with a 6.8% drop in description accuracy, while 2SAC surpasses the baseline’s quality ceiling, improving title accuracy by 5.5% at 5× the throughput and 6× lower latency on identical hardware. We conclude that separating expensive reasoning from real-time serving is both effective and efficient for scalable explanation generation. It eliminates small models’ tendency to editorialize by constraining them to mechanical assembly, and ensures that any product-critical quality dimension is reliably met by a structural property of the pipeline rather than merely a guideline in the prompt.

Authors: Hemanth Ramia Jegdish, Andrés Ferraro and Martín Rocamora

Abstract: Music streaming platforms’ recommender systems have increasingly influenced what users are exposed to. While these recommender systems have been historically designed to maximise user engagement based on click-through rates and other short-term metrics, recent work suggests that increasing diversity can result in long-term retention and conversion. However, diversity in music recommendations can be measured in multiple ways and existing research on this topic lacks reproducibility across datasets and evaluation settings. To address this limitation, we reproduce and extend previous work using a public dataset to investigate how different embedding spaces may affect diversity. We investigate the effect of increasing diversity in a pure collaborative filtering recommender by incorporating more diverse recommendations. We then study how this effect varies when diversity is measured either in a behavioural co-occurrence space or in an acoustic-based embedding space. Our findings show similar diversity–accuracy trade-offs to those reported in previous work when measuring diversity based on a collaborative filtering-based diversity metric, and a stronger effect when diversity is measured in an acoustic embedding space. Additionally, our work provides an evaluation suite to better measure diversity in an offline setting that can be applied to other recommendation tasks or used in future work.