A Deep Matrix Factorization Method for Learning Attribute Representations.

Area of Science:

• Machine Learning
• Data Mining
• Computational Statistics

Background:

• Semi-Non-negative Matrix Factorization (Semi-NMF) offers low-dimensional data representations for clustering.
• Classical clustering methods struggle with complex hierarchical information and implicit attributes in data mappings.
• Existing techniques may not fully capture intricate data structures inherent in the learned representations.

Purpose of the Study:

• To introduce Deep Semi-NMF for learning hierarchical hidden representations interpretable for clustering based on unknown attributes.
• To present Deep WSF, a semi-supervised variant accommodating partial prior information for datasets with mixed attribute knowledge.
• To demonstrate the capability of these models in uncovering complex data structures and enhancing interpretability.

Main Methods:

• Development of a novel Deep Semi-NMF model to learn multi-level latent features.
• Introduction of Deep WSF, a semi-supervised extension incorporating prior knowledge for specific attributes.
• Comparative analysis against Semi-NMF and other state-of-the-art methods.

Main Results:

• Deep Semi-NMF and Deep WSF successfully learn low-dimensional representations with interpretable hierarchical structures.
• The proposed models exhibit superior performance in clustering tasks compared to standard Semi-NMF.
• Both models demonstrate enhanced classification capabilities, outperforming existing state-of-the-art methodologies.

Conclusions:

• Deep Semi-NMF and Deep WSF provide advanced methods for uncovering and interpreting complex hierarchical attributes in datasets.
• These models offer improved clustering and classification performance by learning richer data representations.
• The developed algorithms are effective for datasets with both known and unknown attributes, advancing unsupervised and semi-supervised learning.