SIMCA Modeling for Overlapping Classes: Fixed or Optimized Decision Threshold?

Area of Science:

• Chemometrics
• Machine Learning
• Data Analysis

Background:

• Soft Independent Modeling of Class Analogy (SIMCA) is a widely used classification method.
• Optimizing model complexity and decision thresholds is crucial for SIMCA performance.
• Existing methods may struggle with overlapping sample categories.

Purpose of the Study:

• To propose a novel approach for simultaneous optimization of complexity and decision threshold in SIMCA.
• To leverage Receiver Operating Characteristic (ROC) curve principles for enhanced SIMCA classification.
• To evaluate the proposed method's effectiveness across simulated and real-world datasets.

Main Methods:

• Developed a SIMCA classification approach integrating ROC curve principles.
• Simultaneously optimized model complexity and decision threshold.
• Validated the method using two simulated and four real case-studies.

Main Results:

• The proposed method demonstrated improved classification efficiency in external validation for datasets with strong category overlap.
• The approach enhances robustness against class dispersion compared to fixed threshold methods.
• Satisfactory classification performance was observed for test samples even with clear class separation.

Conclusions:

• The ROC-based optimization offers a more robust and efficient approach to SIMCA classification, particularly for challenging datasets.
• This method provides flexibility in balancing model complexity and classification accuracy.
• The findings suggest broader applicability of ROC principles in chemometric modeling.