The model adaptive space shrinkage (MASS) approach: a new method for simultaneous variable selection and outlier...

Area of Science:

• Chemometrics
• Data mining
• Predictive modeling

Background:

• Variable selection and outlier detection are crucial in chemical modeling, often performed separately.
• The order of these processes significantly impacts modeling outcomes and interpretation.
• Existing methods struggle when prediction errors are similar across different orders.

Purpose of the Study:

• To investigate the interaction between outliers and variables in chemical modeling.
• To compare the effects of different orders of variable selection and outlier detection.
• To develop a simultaneous approach for variable selection and outlier detection.

Main Methods:

• Developed Model Adaptive Space Shrinkage (MASS), a simultaneous variable selection and outlier detection method.
• Utilized model population analysis (MPA) and weighted binary matrix sampling (WBMS).
• Iteratively refined variable and sample weights until convergence to identify optimal subsets.

Main Results:

• MASS adaptively identifies high-performance models with optimized variable and sample subsets.
• The approach effectively cleans datasets for chemical modeling.
• Tested successfully on Near Infrared Spectroscopy (NIR) and Quantitative Structure-Activity Relationship (QSAR) datasets.

Conclusions:

• MASS provides a robust solution by integrating variable selection and outlier detection.
• It overcomes the limitations of sequential, order-dependent methods.
• MASS is a valuable tool for data preprocessing in predictive chemical modeling.