Visualizing Ensemble Predictions of Music Mood.

Area of Science:

• Music Information Retrieval
• Machine Learning
• Data Visualization

Background:

• Music mood classification presents unique challenges compared to other music analysis tasks.
• Ensemble machine learning models offer a potential solution for improving classification accuracy.
• Effective visualization is crucial for understanding complex model predictions and uncertainties.

Purpose of the Study:

• To explore the utility of visualization techniques for music mood classification using ensemble machine learning models.
• To introduce and evaluate a novel visualization method, dual-flux ThemeRiver, for analyzing temporal music data.
• To demonstrate how visualizations can aid in model development and music annotation.

Main Methods:

• Ensemble machine learning models were applied to music mood classification.
• Traditional visualization techniques (stacked line graph, ThemeRiver, pixel-based) were employed.
• A new visualization technique, dual-flux ThemeRiver, was developed and implemented.
• Visualizations were used to analyze popular predictions and model uncertainty across temporal music sections.

Main Results:

• Visualization techniques effectively convey popular predictions and uncertainty in music mood classification.
• The proposed dual-flux ThemeRiver allows for easier observation and measurement of popular predictions compared to existing methods.
• Pixel-based visualization and dual-flux ThemeRiver plots support model development workflows.
• Visualizations facilitate the annotation of music using ensemble model predictions.

Conclusions:

• Visualization techniques are valuable tools for understanding and analyzing music mood classification with ensemble models.
• Dual-flux ThemeRiver offers an improved method for visualizing temporal music data and model performance.
• These visualization approaches enhance both the analysis of individual models and the application of ensemble predictions.