Interpretable unsupervised neural network structure for data clustering via differentiable reconstruction of ONMF and...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Neural networks often lack interpretability in clustering tasks.
• Traditional methods rely on post-hoc explanations or supervised learning, limiting transparency.
• Existing deep clustering methods struggle with end-to-end trainability and clear interpretability.

Purpose of the Study:

• To develop a novel neural network model for interpretable clustering.
• To enhance the transparency and performance of clustering tasks through integrated factorization methods.
• To enable end-to-end training of neural networks with embedded clustering layers.

Main Methods:

• Integration of Orthogonal Non-negative Matrix Factorization (ONMF) into a neural network clustering layer.
• Enhancement of the model using a Sparse Autoencoder (SAE) for feature extraction.
• Development of a differentiable ONMF reconstruction for end-to-end training and interpretability.

Main Results:

• OSINN achieves high clustering accuracies on benchmark datasets: MNIST (90%), CIFAR-10 (24%), Fashion-MNIST (64%), and CIFAR-100 (44%).
• The model demonstrates significant performance improvements over traditional clustering algorithms (+10%) and deep clustering methods (+1.2%).
• OSINN surpasses Non-negative Matrix Factorization (NMF) and Autoencoder (AE) variants by over 1.5% in clustering performance.

Conclusions:

• OSINN offers a transparent and interpretable approach to neural network-based clustering.
• The end-to-end trainable architecture enhances both performance and understanding of clustering results.
• This method is particularly effective for large, unlabeled datasets, advancing unsupervised learning capabilities.