An Interpretable Adaptive Multiscale Attention Deep Neural Network for Tabular Data

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning (DL) excels at analyzing signals like sound and images.
• DL faces performance limitations with tabular structured data compared to shallow learning.
• Explainability remains a challenge in deep learning models for tabular data.

Purpose of the Study:

• To introduce a novel deep learning architecture for tabular data.
• To improve the performance of deep learning on classification and regression tasks using tabular data.
• To enhance the explainability of deep learning models applied to tabular data.

Main Methods:

• Developed an adaptive multiscale attention deep neural network architecture.
• Employed parallel multilevel feature weighting for enhanced pattern recognition.
• Implemented four distinct levels of explainability for model comprehension.

Main Results:

• Achieved high F1-scores on seven classification tasks across various dataset sizes.
• Obtained low mean absolute errors on four regression tasks.
• Demonstrated the model's ability to learn feature attention and provide interpretable insights.

Conclusions:

• Adaptive multiscale attention significantly improves deep learning performance on tabular data.
• The architecture offers unprecedented explainability, aiding in understanding model behavior.
• This technique is valuable for feature ranking and selection in tabular datasets.