Deep Multirepresentation Learning for Data Clustering.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Conventional deep clustering methods utilize a single global embedding subspace for all data clusters.
• This approach can be suboptimal for complex datasets with varying cluster characteristics.

Purpose of the Study:

• To propose a novel deep multirepresentation learning (DML) framework for data clustering.
• To address limitations of single-latent-space methods, particularly for imbalanced datasets.

Main Methods:

• Developed a DML framework employing autoencoders (AEs) to generate distinct cluster-specific and a general latent space.
• Introduced a specialized loss function weighting reconstruction and clustering losses based on sample probability for cluster assignment.
• Utilized weighted reconstruction and clustering losses for AE specialization.

Main Results:

• The proposed DML framework and loss function demonstrated superior performance over state-of-the-art clustering approaches on benchmark datasets.
• DML significantly outperformed existing methods on imbalanced datasets by assigning individual latent spaces to challenging clusters.

Conclusions:

• The DML framework offers a more effective approach to deep clustering by adapting latent spaces to data complexity.
• This method shows particular promise for improving clustering performance on datasets with significant class imbalance.