Learning Markov Random Walks for robust subspace clustering and estimation.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Conventional Markov Random Walks (MRW) struggle with data from mixed subspaces due to limited global structural measurement.
• Existing methods lack the ability to effectively handle complex subspace structures.

Purpose of the Study:

• To develop a regularized MRW learning model for robust subspace clustering and estimation.
• To integrate local pairwise similarity and global subspace structure learning within the MRW framework.
• To enhance the model's performance on data with multiple subspaces.

Main Methods:

• Introduced a regularized MRW learning model with a low-rank penalty to constrain global subspace structure.
• Learned local and global structures from MRW transition probabilities.
• Proposed a robust extension integrating transition matrix learning and error correction.

Main Results:

• The proposed model accurately captures multiple subspace structures and learns low-dimensional embeddings.
• Demonstrated exact segmentation of subspaces under specific conditions.
• The robust extension improved performance in real-world scenarios.

Conclusions:

• The regularized MRW learning model offers superior performance for subspace clustering compared to state-of-the-art methods.
• The model effectively handles data with mixed subspace distributions.
• The proposed framework provides a robust and accurate approach to subspace analysis.