Evolving fuzzy neural classifier that integrates uncertainty from human-expert feedback

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining

Background:

• Data quality significantly impacts model performance.
• Labeling uncertainty in data can arise from expert inexperience or low confidence.
• Evolving fuzzy neural networks (EFNC) are powerful tools for complex problem-solving.

Purpose of the Study:

• To introduce EFNC-U, an approach integrating expert labeling uncertainty into EFNC.
• To enhance the interpretability of fuzzy classification rules.
• To improve the accuracy and robustness of EFNC models when dealing with uncertain data.

Main Methods:

• Proposed EFNC-U by incorporating expert input on labeling uncertainty.
• Conducted binary pattern classification tests in cyber invasion and auction fraud detection scenarios.
• Evaluated model accuracy and rule interpretability with varying levels of simulated uncertainty.

Main Results:

• EFNC-U demonstrated improved accuracy trends compared to models trained on uncertain data without explicit uncertainty consideration.
• The approach showed robustness, with accuracy trends similar to original data streams for uncertainty levels below 20%.
• Generated interpretable fuzzy classification rules with reduced antecedent lengths and certainty values in consequent labels.

Conclusions:

• Integrating expert labeling uncertainty into EFNC models is effective for improving performance and interpretability.
• EFNC-U offers a robust solution for handling uncertain data in classification tasks.
• The elicited rules provide valuable insights and knowledge discovery potential for applications like fraud detection.