UnAJoEmGMR: Uncertainty-aware joint emotion and genre classification for context-aware music recommendation
Year: 2026
Type: article
Abstract: A B S T R A C T Music is an important aspect of daily life that influences mood, motivation, and behavioral context, making music recommender systems an effective means of improving user experiences. This research presents an uncertainty-aware joint emotion and genre classification framework for context-aware music recommendation (UnAJoEmGMR) that suggests audio tracks based on the user’s emotional state and physical activity context. Two task-specific convolutional neural network branches process spectrogram representations to learn emotion- and genre-related features through a cross-task attention mechanism. The proposed architecture incorporates a Bayesian approximation by enabling Monte Carlo dropout during inference, allowing predictive uncertainty to be estimated. Unlike conventional approaches that rely on user metadata or listening history, the proposed system maps inferred genre information to physical activity levels, enabling context-aware music recommendation. On average, the proposed classification framework showed 89.39 ± 3.73% for emotions and 89.51 ± 3.65% for genre among the 10 independent tests. The recommendation framework was assessed using ranking-based metrics to evaluate the suitability of tracks. These points concerned the listeners’ emotional state and activity. At K = 50, 100, and 150, the Precision@K scores were 0.86, 0.85, and 0.77, respectively. At all three points of evaluation, the NDCG scores were over 0.97; exhibiting evidence for the context-aware music recommendation approach.
Source: Array
Authors: K.P. Faseela, P. Supriya
Institution: Amrita Vishwa Vidyapeetham
Language: English
Topic: Music and Audio Processing
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,680
OpenAlex ID: W7203740609