Semi-supervised deep embedded clustering with pairwise constraints and subset allocation.

Area of Science:

• Machine Learning
• Data Mining
• Natural Language Processing

Background:

• Semi-supervised deep clustering methods are effective for end-to-end clustering.
• Overlapping samples in industrial text datasets negatively impact clustering performance.
• Existing methods struggle with integrating pairwise constraints and class labels effectively.

Purpose of the Study:

• To propose a novel semi-supervised deep clustering method (PCSA-DEC) for industrial text datasets.
• To address the limitations of existing methods in handling overlapping samples and integrating supervision information.
• To improve clustering accuracy and normalized mutual information.

Main Methods:

• Developed a semi-supervised method based on pairwise constraints and subset allocation (PCSA-DEC).
• Redefined similarity-based constraint loss to enhance intra-class similarity.
• Designed a novel subset allocation loss for precise learning of strong-supervised information.

Main Results:

• PCSA-DEC achieved 8.2%-8.7% improvement in accuracy on industrial text datasets.
• The method showed a 13.4%-19.8% improvement in normalized mutual information compared to state-of-the-art methods.
• Demonstrated superior performance in handling overlapping samples and integrating supervision.

Conclusions:

• The proposed PCSA-DEC method effectively improves semi-supervised deep clustering for industrial text.
• The novel loss functions successfully address challenges posed by overlapping samples and weak/incorrect supervision.
• PCSA-DEC offers a robust solution for accurate and reliable text clustering in industrial applications.