Unsupervised feature selection via latent representation learning and manifold regularization.

Area of Science:

• Machine Learning
• Data Mining
• Multimedia Technology

Background:

• Massive unlabelled, high-dimensional data present processing challenges.
• Traditional unsupervised feature selection methods struggle with data dependencies and noise.
• Existing methods often fail in the original data space due to noise interference.

Purpose of the Study:

• To propose a robust unsupervised feature selection method for high-dimensional, unlabelled data.
• To embed latent representation learning into feature selection for improved noise robustness.
• To leverage instance relationships for effective feature selection without label information.

Main Methods:

• Feature selection performed in a learned latent representation space, robust to noise.
• Latent representation modeled using non-negative matrix factorization of the affinity matrix.
• Graph-based manifold regularization preserves local data structure in the transformed space.

Main Results:

• The proposed method demonstrates effectiveness across eight benchmark datasets.
• Feature selection in the latent space proves more robust to noise than original data space methods.
• Experimental results validate the method's ability to handle interconnected data instances.

Conclusions:

• The developed method offers a robust approach to unsupervised feature selection for complex datasets.
• Integrating latent representation learning enhances feature selection performance and noise resilience.
• The approach effectively utilizes instance relationships and manifold structures for relevant feature identification.