Semantic IDs for Music Recommendation

2025-07-24

AI Summary

To address the inefficiency and poor generalization of one-hot ID embeddings in music recommendation—characterized by excessive parameter count and limited semantic expressiveness—this paper proposes Semantic ID, a content-aware shared embedding framework. Semantic ID learns low-dimensional, semantically meaningful representations from item content features via end-to-end joint optimization of a content encoder and the recommendation model, thereby drastically reducing the number of independent ID embeddings. Evaluated on two public music datasets, it outperforms baseline methods across key metrics including Recall@10 and Intra-List Similarity (ILS), while reducing model parameters by 37%–52%. Moreover, it improves both recommendation accuracy and diversity. Online A/B testing in a real-world streaming service confirms its operational effectiveness and delivers measurable business gains.

Technology Category

• Data Mining & Knowledge Management: Recommender Systems
• Machine Learning: Learning Preferences or Rankings
• Search and Optimization: Learning to Search

Application Category

• User Modeling, Personalization and Recommendation: Personalized content ranking and presentation
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs

Abstract

Training recommender systems for next-item recommendation often requires unique embeddings to be learned for each item, which may take up most of the trainable parameters for a model. Shared embeddings, such as using content information, can reduce the number of distinct embeddings to be stored in memory. This allows for a more lightweight model; correspondingly, model complexity can be increased due to having fewer embeddings to store in memory. We show the benefit of using shared content-based features ('semantic IDs') in improving recommendation accuracy and diversity, while reducing model size, for two music recommendation datasets, including an online A/B test on a music streaming service.

Problem

• Reducing memory usage with shared embeddings
• Improving music recommendation accuracy and diversity
• Testing semantic IDs in streaming service A/B

Innovation

• Shared content-based semantic IDs reduce embeddings
• Lightweight model with increased complexity potential
• Improves accuracy and diversity in recommendations