A Data-Driven Framework for Class Overlap Reduction to Improve Soft Sensor Performance

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Class overlap and imbalance degrade machine learning performance in data-driven applications.
• Existing resampling methods often fail to preserve majority class distribution after cleaning.
• Addressing these issues independently limits effectiveness.

Purpose of the Study:

• Propose a novel framework, Two-Phase Clustering Plus, to effectively handle class overlap and imbalance.
• Develop a method that preserves the underlying distribution of the majority class.
• Improve machine learning performance on complex imbalanced datasets.

Main Methods:

• Leverage overlap detection to guide the undersampling process.
• Integrate Edited Nearest Neighbors to isolate overlapping regions for data cleaning.
• Apply a two-phase clustering strategy to retain safe majority data and preserve boundary samples.

Main Results:

• Achieved a 1.45% improvement in geometric mean on service industry datasets.
• Significantly enhanced class separability.
• Demonstrated a robust and distribution-aware solution for imbalanced datasets.

Conclusions:

• Two-Phase Clustering Plus offers a targeted elimination of ambiguous majority instances without significant information loss.
• The framework systematically selects representative samples, maintaining critical boundaries and global data structure.
• This approach provides a superior solution for complex imbalanced datasets in virtual sensing environments.