Aligning Music Recommendations with Human Moods using Graph Learning

What is it about?

Recommender systems for music typically rely on Graph Contrastive Learning (GCL) to predict user preferences based on past interaction histories. However, standard graph models treat all user-item interactions as identical structural links, ignoring the emotional tone of content. This causes "semantic incongruence", for instance, treating an upbeat dance anthem and a melancholic ballad as similar simply because a user listened to both in the same session. To resolve this limitation without adding heavy neural network complexity, we introduce ACE (Affective Contrastive Embeddings). ACE extracts psychological dimensions of emotion, valence (positivity), and arousal (energy level) from Large Language Models (LLMs) to create a continuous 2D affective manifold. Rather than building complex auxiliary neural network layers, ACE enforces a geometric constraint called Semantic Isomorphism directly within the loss landscape using a novel objective function (Semantically Weighted Noise Contrastive Estimation). This enables the model to automatically suppress gradient signals from emotionally dissonant interactions while accelerating convergence for emotionally resonant items. Additionally, ACE incorporates user-specific adaptive sensitivity coefficients to distinguish mood-sensitive listeners from eclectic explorers.

Featured Image

Photo by C D-X on Unsplash

Photo by C D-X on Unsplash

Why is it important?

Geometric Constraint vs. Architectural Complexity: Existing context-aware recommendation systems rely on "architectural stacking", layering complex attention heads or auxiliary encoders onto graph backbones. ACE proves that context can be effectively integrated directly into the loss landscape, allowing the model to operate as a lightweight "0-layer" network during inference. Sustainable & Green AI: By shifting structural aggregation from forward propagation to backward pass optimization, ACE achieves superior retrieval accuracy over 11 state-of-the-art baselines across Amazon Digital Music, Last.fm, and SiTunes, while drastically cutting inference latency, VRAM usage, and training carbon emissions. User Heterogeneity & Adaptive Modeling: Rather than forcing rigid mood constraints onto every user, ACE dynamically balances semantic consistency for mood-driven listeners with structural exploration for eclectic listeners. LLMs as Reliable Semantic Oracles: The work validates that ensemble-distilled Large Language Models can act as accurate "Semantic Oracles," bridging the gap between world knowledge and user interaction graphs without manual labeling.

Perspectives