Deep Self-Evolution Clustering.

Area of Science:

• Machine Learning
• Pattern Analysis
• Artificial Intelligence

Background:

• Clustering is a fundamental yet challenging task in pattern analysis and machine learning.
• Current clustering methods often fail to integrate representation learning effectively.
• This limitation hinders the ability to extract meaningful patterns and achieve optimal clustering performance.

Purpose of the Study:

• To develop a novel deep learning framework, Deep Self-Evolution Clustering (DSEC), that jointly learns representations and performs clustering.
• To address the limitations of existing methods by integrating representation learning directly into the clustering process.
• To enhance the accuracy and efficiency of clustering algorithms for complex datasets.

Main Methods:

• DSEC reframes clustering as a binary pairwise-classification problem, estimating pattern similarity.
• It utilizes a deep neural network (DNN) to generate indicator features for pairwise pattern similarity assessment.
• An iterative algorithm, Self-Evolution Clustering Training (SECT), alternately selects similar/dissimilar pairs and trains the DNN, enabling unsupervised learning of representations.

Main Results:

• The proposed DSEC method demonstrates superior performance across twelve diverse datasets, including image, text, and audio data.
• Experiments show consistent improvements over existing state-of-the-art clustering models.
• The learned indicator features generated by DSEC tend towards one-hot vectors, facilitating effective pattern clustering.

Conclusions:

• DSEC offers a significant advancement in clustering by effectively combining representation learning and clustering tasks.
• The Self-Evolution Clustering Training (SECT) algorithm provides a robust mechanism for unsupervised learning of informative representations.
• DSEC establishes a new benchmark for clustering performance, particularly in complex, high-dimensional data scenarios.