Related Experiment Video
Updated: Jan 10, 2026
08:00
Decoding Natural Behavior from Neuroethological Embedding
Published on: October 3, 2025
569
Pseudo datasets estimate feature attribution in artificial neural networks
Hui-Yi Yang1, Yi-Hau Chen2, Hao-Min Cheng3,4,5,6
1Division of Biostatistics and Data Science, Institute of Public Health, College of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan, ROC.
Scientific Reports
|November 22, 2025
Summary
This study introduces Pseudo Datasets Perturbation Effect (PDPE), a new method for explaining neural network predictions. PDPE effectively identifies feature interactions and individual feature importance, offering faster and more accurate insights than existing techniques.
Area of Science:
- Artificial Intelligence
- Machine Learning
- Data Science
Background:
- Neural networks excel at classification but lack interpretability, hindering their application.
- Current explanation methods often overlook feature interactions, limiting a full understanding of model behavior.
Purpose of the Study:
- To propose a novel two-stage explanation method, Pseudo Datasets Perturbation Effect (PDPE), for neural networks.
- To address the limitations of existing methods by accounting for feature interactions and individual feature significance.
Main Methods:
- PDPE perturbs data to observe the impact on predictions, discerning feature importance and interactions.
- Evaluates individual features and their interaction terms within structured data.
Main Results:
- Computer simulations show PDPE is faster and more accurate than SHAP Value for explaining neural networks.
- Real-life data analysis from the National Institute of Diabetes and Digestive and Kidney Diseases confirms PDPE's superior performance.
Conclusions:
- PDPE enhances understanding of how individual features and their interactions influence neural network predictions.
- The method offers a more comprehensive and efficient approach to model interpretability in machine learning.
