Unsupervised Feature Selection via Adaptive Multimeasure Fusion.

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• Estimating data point similarity often involves diverse criteria, leading to varied pairwise relation representations.
• Existing methods struggle with integrating multiple similarity measures effectively.

Purpose of the Study:

• To propose a novel self-adaptive multimeasure (SAMM) fusion approach for unifying diverse similarity measures.
• To develop an adaptive similarity evaluation method that optimizes similarity as a variable.
• To integrate graph-based dimensionality reduction and sparsity-inducing regularization for subspace representation and feature selection.

Main Methods:

• Developed a self-adaptive multimeasure (SAMM) fusion framework.
• Incorporated a graph-based dimensionality reduction technique.
• Introduced sparsity-inducing l_2,0 regularization for feature selection (FS).

Main Results:

• The SAMM approach adaptively merges different measure functions into a unified similarity measure.
• The integrated method achieves subspace representation based on the unified similarity.
• The SAMM-FS method yields a sparse projection for efficient feature selection.

Conclusions:

• The proposed SAMM-FS method effectively addresses the challenge of diverse similarity representations.
• This adaptive fusion approach enhances the accuracy and efficiency of similarity estimation and feature selection.
• The method offers a robust solution for data analysis tasks requiring integrated similarity measures.