Entropy Minimizing Matrix Factorization.

Area of Science:

• Data Science
• Machine Learning
• Matrix Factorization

Background:

• Nonnegative matrix factorization (NMF) is a common technique for data analysis.
• Traditional NMF methods are sensitive to outliers, which can skew results.
• Outliers can dominate the objective value in standard NMF, affecting normal sample approximation.

Purpose of the Study:

• To develop a novel framework, entropy minimizing matrix factorization (EMMF), to mitigate the impact of outliers in NMF.
• To improve the robustness and accuracy of matrix factorization techniques when dealing with noisy datasets.
• To enhance data analysis by ensuring that outliers do not disproportionately influence the model's objective.

Main Methods:

• Developed a new entropy loss function for matrix factorization that minimizes residue distribution entropy.
• Derived multiplicative updating rules for the EMMF framework and provided theoretical convergence proofs.
• Introduced a Graph regularized version of EMMF (G-EMMF) incorporating data graph relationships.

Main Results:

• EMMF framework effectively minimizes the impact of outliers on the objective value.
• The proposed EMMF and G-EMMF models demonstrate superior performance in clustering tasks.
• Experimental results on synthetic and real-world datasets validate the effectiveness against state-of-the-art methods.

Conclusions:

• The entropy minimizing matrix factorization (EMMF) framework offers a robust solution for handling outliers in nonnegative matrix factorization.
• The G-EMMF variant further enhances performance by leveraging data relationships through graph regularization.
• EMMF provides improved accuracy and reliability for data analysis tasks, particularly in the presence of outliers.