Autoencoder Constrained Clustering With Adaptive Neighbors.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Conventional subspace clustering methods rely on explicit data representation and struggle with intrinsic linearity and fixed structures.
• Existing methods have limitations in capturing complex, nonlinear data relationships and adapting to evolving data patterns.

Purpose of the Study:

• To develop an adaptive deep clustering approach that overcomes the limitations of conventional subspace clustering.
• To integrate structured graph learning with adaptive neighbors into deep autoencoder networks for enhanced clustering.
• To enable adaptive investigation of nonlinear data structures and iterative strengthening of deep feature correlations.

Main Methods:

• Developed an adaptive deep clustering approach named autoencoder constrained clustering with adaptive neighbors (ACC_AN).
• Embedded structured graph learning with adaptive neighbors into deep autoencoder networks.
• Utilized a parameter-free graph built upon deep features to investigate nonlinear data structures.
• Minimized reconstruction error to preserve the local structure of raw data.

Main Results:

• ACC_AN adaptively investigates nonlinear data structures using a parameter-free graph.
• The method iteratively strengthens correlations among deep representations during the learning process.
• Preserves local data structure by minimizing reconstruction error, outperforming state-of-the-art methods.

Conclusions:

• ACC_AN is the first deep clustering method to embed adaptive structured graph learning for simultaneous updates of latent data representation and the structured deep graph.
• The proposed method offers a novel approach to adaptive deep clustering by effectively handling nonlinearities and preserving local structures.