Safe Classification with Augmented Features.

Area of Science:

• Machine Learning
• Data Science
• Computational Psychiatry

Background:

• Increasingly, data collection yields abundant features, but more features do not guarantee improved classification performance.
• Preventing performance degradation from augmented features is critical yet understudied in classification tasks.

Purpose of the Study:

• To propose a novel safe classification approach (SAC) that guarantees classification accuracy is never degenerated when using augmented features.
• To address the challenge of feature augmentation potentially worsening classification outcomes.

Main Methods:

• Developed SAfe Classification (SAC) using two key strategies: learning diverse classifiers with a robust loss function and integrating classifiers for a safe prediction.
• Proposed new optimization methods with proven convergence to handle the computational challenges.
• Theoretically guaranteed safeness under mild assumptions.

Main Results:

• Evaluated SAC on 16 diverse datasets, demonstrating its effectiveness in handling augmented features without performance loss.
• Successfully applied SAC to the diagnostic classification of schizophrenia, showing its practical utility.
• Experimental results confirmed SAC's ability to discriminate schizophrenic patients from healthy controls.

Conclusions:

• SAC provides a theoretically sound and practically effective method for safe feature augmentation in classification.
• The approach holds significant potential for applications like medical diagnosis, particularly in distinguishing complex conditions such as schizophrenia.