Predicting danceability and song ratings using deep learning and auditory features
Wu Wei · PeerJ Computer Science · 2025
Predicting a song's danceability and overall rating poses a significant challenge due to the complex interplay between musical characteristics and listener preferences. In this study, we propose a deep learning framework that jointly addresses the tasks of danceability estimation and popularity prediction. Our model integrates a Bidirectional Long Short-Term Memory (BiLSTM) network to capture sequential and contextual patterns from categorical inputs, alongside a Residual Network (ResNet) that extracts hierarchical representations from numerical auditory features. These complementary feature streams are fused using a cross-attention mechanism, enabling the model to effectively learn intricate relationships across heterogeneous data modalities. Experimental evaluations demonstrate that our approach consistently outperforms traditional machine learning baselines and recent deep learning models. The results demonstrate the effectiveness of cross-attention in structured music data modelling and highlight the framework's potential in advancing music recommendation and audio analysis systems.