Local Adaptive Projection Framework for Feature Selection of Labeled and Unlabeled Data

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• Traditional feature selection methods rely on fixed similarity matrices, which can be unreliable due to noise and may fail to capture local data structures.
• Noise features can significantly impact the accuracy of similarity computations, leading to suboptimal feature selection outcomes.
• Existing methods often struggle to preserve the intricate local structure within data classes when computing similarities.

Purpose of the Study:

• To propose a novel local adaptive projection (LAP) framework for robust feature selection.
• To develop a method that simultaneously learns adaptive similarity and projection matrices.
• To enhance feature selection by effectively addressing noise and preserving local data structures.

Main Methods:

• The proposed Local Adaptive Projection (LAP) framework iteratively learns an adaptive similarity matrix and a projection matrix.
• Similarity is computed from projected distances, and the projection matrix is derived from the learned similarity in each iteration.
• Two specific methods, Supervised Feature Selection with LAP (SLAP) and Unsupervised Feature Selection with LAP (ULAP), are introduced.

Main Results:

• Experimental results demonstrate the effectiveness of the LAP framework across eight diverse datasets.
• SLAP significantly outperformed seven existing supervised feature selection methods.
• ULAP showed superiority over five established unsupervised feature selection methods.

Conclusions:

• The LAP framework offers a more robust and adaptive approach to feature selection compared to traditional methods.
• By adaptively learning similarity and projection, LAP effectively reduces the impact of noise features.
• The proposed SLAP and ULAP methods provide superior performance in both supervised and unsupervised feature selection tasks.