Adaptive Feature Selection With Augmented Attributes.

Area of Science:

• Machine Learning
• Data Science
• Neuroimaging Analysis

Background:

• Dynamic environments generate incremental data with accumulating feature spaces, posing challenges for high-dimensional data manipulation.
• Neuroimaging applications, like diagnosing neuropsychiatric disorders, face difficulties in feature selection due to evolving data and diverse testing methods.

Purpose of the Study:

• To propose a novel Adaptive Feature Selection (AFS) method for feature incremental scenarios.
• To enable reusability of previously trained feature selection models and adapt them to new feature sets.
• To address the challenge of selecting valuable features in high-dimensional, evolving datasets.

Main Methods:

• Developed an Adaptive Feature Selection (AFS) method incorporating an l0-norm sparse constraint.
• Proposed an effective solving strategy for the l0-norm constraint in feature selection.
• Extended the method from a one-shot case to a multi-shot scenario, analyzing generalization bounds and convergence.

Main Results:

• Experimental results demonstrate the effectiveness of reusing previous features in feature selection.
• The l0-norm constraint proved superior in various aspects compared to other methods.
• The AFS method showed effectiveness in discriminating schizophrenic patients from healthy controls.

Conclusions:

• The proposed Adaptive Feature Selection (AFS) method effectively handles incremental feature selection in dynamic environments.
• Reusing previously selected features and employing an l0-norm constraint significantly enhance performance.
• AFS shows promise for applications in neuroimaging and other fields dealing with high-dimensional, evolving data.