Adaptive Attribute and Structure Subspace Clustering Network.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Deep self-expressiveness methods are effective for subspace clustering.
• Existing methods often limit performance by only using attribute information for self-expression.

Purpose of the Study:

• To propose a novel adaptive attribute and structure subspace clustering network (AASSC-Net).
• To simultaneously consider attribute and structure information using adaptive graph fusion.

Main Methods:

• Utilized an auto-encoder for latent feature representation and attribute matrix construction.
• Constructed a mixed signed and symmetric structure matrix to capture local geometric data structure.
• Employed self-expressiveness on attribute and structure matrices to learn separate affinity graphs.
• Developed an attention-based fusion module for adaptive graph combination.

Main Results:

• AASSC-Net significantly outperforms state-of-the-art methods on benchmark datasets.
• Ablation studies confirm the effectiveness of the designed modules.
• The proposed method achieves superior discriminative affinity graph construction.

Conclusions:

• The proposed AASSC-Net effectively integrates attribute and structure information for enhanced subspace clustering.
• Adaptive graph fusion is a key factor in improving clustering performance.
• The method offers a significant advancement in subspace clustering techniques.