Predicting danceability and song ratings using deep learning and auditory features.

Area of Science:

• Music Information Retrieval
• Deep Learning
• Computational Musicology

Background:

• Predicting song danceability and popularity is challenging due to complex musical features and listener preferences.
• Existing models struggle to integrate diverse musical data effectively.

Purpose of the Study:

• To develop a deep learning framework for joint danceability estimation and popularity prediction.
• To leverage heterogeneous data modalities for improved music analysis.

Main Methods:

• Utilized a Bidirectional Long Short-Term Memory (BiLSTM) network for sequential categorical data.
• Employed a Residual Network (ResNet) for hierarchical numerical auditory features.
• Integrated feature streams using a cross-attention mechanism for multimodal data fusion.

Main Results:

• The proposed framework demonstrated superior performance compared to traditional machine learning and recent deep learning models.
• Cross-attention mechanism proved effective in modeling structured music data.
• The model successfully learned intricate relationships across heterogeneous data.

Conclusions:

• The developed deep learning framework offers a robust solution for music data modeling.
• The approach shows significant potential for enhancing music recommendation systems and audio analysis.
• Cross-attention is a key mechanism for integrating diverse musical features.