Attribution-based interpretable classification neural network with global and local perspectives.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Explainable AI (XAI)

Background:

• Neural networks' black-box nature hinders reliability in critical domains.
• Existing interpretability methods often compromise model performance or provide only local explanations.

Purpose of the Study:

• To develop an attribution-based interpretable classification model for tabular data.
• To achieve both high classification performance and detailed model interpretability.

Main Methods:

• Mapping intermediate neural network outputs to an interpretable data representation space.
• Developing an attribution-based model that selects relevant features for classification and interpretation.
• Investigating training strategies to balance performance and interpretability.

Main Results:

• The proposed model demonstrates classification accuracy comparable to black-box neural networks on eight datasets.
• The model provides both local and global feature importance, enhancing interpretability.
• Outperformed popular post-hoc interpretability methods on Reverse Precision and Generality metrics.

Conclusions:

• The developed model offers a viable solution for interpretable classification on tabular data.
• Highlights a trade-off between classification performance and interpretability during model training.
• Achieves competitive accuracy while significantly improving model transparency.