Deep Tensor Spectral Clustering Network via Ensemble of Multiple Affinity Tensors

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Tensor spectral clustering (TSC) leverages multi-wise similarities for enhanced learning.
• Existing TSC methods struggle with memory-intensive high-order affinity tensors and suboptimal two-stage learning processes.

Purpose of the Study:

• To propose a novel Tensor Spectral Clustering Network (TSC-Net) for efficient and accurate clustering.
• To address the memory and performance limitations of current TSC approaches.

Main Methods:

• TSC-Net employs a deep neural network for one-stage learning of a consensus tensor spectral embedding.
• Stochastic optimization is utilized to compute affinity tensors incrementally, significantly reducing memory footprint.
• An ensemble of multiple affinity tensors guides the TSC objective within the network.

Main Results:

• TSC-Net achieves significant memory cost reduction compared to existing methods.
• The one-stage learning approach leads to improved clustering performance.
• Empirical evaluations on benchmark datasets show TSC-Net outperforms recent baseline methods.

Conclusions:

• TSC-Net offers an efficient and effective solution for tensor spectral clustering.
• The proposed method overcomes key challenges in memory usage and clustering accuracy.
• TSC-Net represents a significant advancement in applying TSC for complex data analysis.