Truncated Cauchy Non-Negative Matrix Factorization.

Area of Science:

• Machine Learning
• Data Science
• Robust Statistics

Background:

• Non-negative Matrix Factorization (NMF) is sensitive to outliers due to its reliance on Euclidean distance.
• Corrupted data can lead to inaccurate low-rank approximations in standard NMF.

Purpose of the Study:

• To develop a robust Non-negative Matrix Factorization method capable of handling datasets with outliers.
• To introduce a novel Truncated CauchyNMF loss function for improved subspace learning on noisy data.

Main Methods:

• Proposed a Truncated CauchyNMF loss function that mitigates the impact of large errors by truncation.
• Developed the Truncated CauchyNMF algorithm for robust subspace learning.
• Theoretically analyzed the robustness and generalization bounds of Truncated CauchyNMF, proving convergence rates.

Main Results:

• Truncated CauchyNMF demonstrates superior robustness compared to competing NMF models on corrupted datasets.
• Theoretical analysis confirmed a generalization bound for Truncated CauchyNMF with a convergence rate of order O(1/sqrt(n)).
• Image clustering experiments on simulated and real-world noisy datasets validated the effectiveness of Truncated CauchyNMF.

Conclusions:

• Truncated CauchyNMF is an effective and robust method for learning subspaces from noisy data contaminated by outliers.
• The proposed loss function and algorithm provide a significant advancement for NMF applications in the presence of data corruption.
• The theoretical guarantees support the practical performance observed in experiments.