Methods
Iterative feature boosting with attribution
Using Shapley-based explanations to examine acoustic features and refine an emotion-recognition model.
Feature attribution can be part of model development rather than only a report produced after training. The iterative feature-boosting studies use Shapley-based analysis to examine how acoustic features contribute to predictions and to guide changes to the feature set.
This method connects interpretability with an experimental loop. The conference and journal studies have distinct publication records and evaluation scopes; their linked papers provide the full procedure and results.
Walk through the method
Open each step to inspect the inputs, process and outputs described by its authors.
01 Establish a feature-based model
- Input
- Acoustic features and emotion labels
- Output
- Initial predictive model
Start from acoustic descriptors and a speech-emotion prediction task. The feature set defines what information the model is able to use and provides the starting point for attribution analysis.
02 Examine feature contributions
- Input
- Model and feature set
- Output
- Feature-attribution evidence
Shapley-based analysis helps identify which features contribute to the model’s predictions. The studies examine relevance and redundancy to inform the next iteration of the representation.
03 Refine and evaluate
- Input
- Attribution-guided feature choices
- Output
- Revised model and comparative evaluation
Use the attribution analysis to refine the feature set, then evaluate the resulting model. Inspect predictive results together with what changed in the representation, following the protocol in the corresponding paper.
Projects
- 01
Projects
Explainable feature boosting
Using feature attribution to refine acoustic representations and examine model decisions.
Research directions
- 01
Research programs
Interpretable & multimodal learning
Understanding acoustic features and examining how audio, language and visual cues interact in emotion recognition.