Integrating Tensor Similarity to Enhance Clustering Performance.

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• Clustering performance relies heavily on pairwise similarity matrices.
• Pairwise similarity is susceptible to noise and data imbalance, limiting clustering accuracy.

Purpose of the Study:

• To propose a novel clustering approach that leverages higher-order sample information.
• To enhance clustering robustness against noise and data imbalance.

Main Methods:

• A fourth-order tensor similarity is derived, representing higher-order relationships among samples.
• Eigenvalue decomposition is used to obtain a high-order similarity matrix from the tensor.
• The high-order similarity is integrated with pairwise similarity to form the IPS2 method.

Main Results:

• The proposed high-order similarity captures spatial information, complementing pairwise similarity.
• IPS2 significantly outperforms existing similarity-based clustering methods on real-world datasets.
• IPS2 demonstrates capability in handling under-sampled and noisy datasets.

Conclusions:

• Integrating higher-order sample information offers a robust solution for clustering challenges.
• IPS2 provides a significant advancement in clustering performance, particularly for noisy and imbalanced data.