Robust dimensionality reduction via feature space to feature space distance metric learning.

Area of Science:

• Computer Science
• Machine Learning
• Data Science

Background:

• High-dimensional data in image classification poses challenges like the curse of dimensionality.
• Existing methods like Laplacian Eigenmaps (LE) struggle with noise robustness.
• Point-to-point (P2P) distance metrics in LE lack sufficient robustness.

Purpose of the Study:

• To develop a novel supervised dimensionality reduction method for improved classification.
• To enhance robustness to noise in high-dimensional data.
• To better mine local geometric information for classification tasks.

Main Methods:

• Introduced Feature Space to Feature Space Distance Metric Learning (FSDML).
• Constructed feature spaces using k-intra-class nearest neighbors for local projections.
• Defined a Space-to-Space (S2S) distance metric based on Euclidean distance between projections.
• Modeled intra-class and inter-class graphs using label similarity and dissimilarity.
• Optimized classification by maximizing S2S manifold distances and preserving locality.

Main Results:

• The proposed S2S distance metric demonstrated superior robustness to noise.
• Local projections effectively captured local geometry information from the data.
• Experimental results on synthesized and benchmark datasets validated FSDML's performance.
• FSDML outperformed state-of-the-art dimensionality reduction methods.

Conclusions:

• FSDML offers a robust and effective approach to supervised dimensionality reduction.
• The method successfully addresses limitations of traditional LE techniques.
• FSDML shows significant potential for improving classification accuracy in high-dimensional image data.