A lightweight mixup-based short texts clustering for contrastive learning.

Area of Science:

• Computational linguistics
• Medical informatics
• Machine learning

Background:

• Traditional text clustering methods face challenges with overlapping representations in medical data.
• Clustering is crucial for categorizing medical conditions from large unlabeled text datasets.
• Clustering sparse data, like that for rare diseases, presents unique difficulties.

Purpose of the Study:

• To propose a novel contrastive clustering method enhanced with mixup for medical text analysis.
• To address the limitations of traditional clustering in handling overlapping data and sparse rare disease information.
• To optimize feature spaces and reduce computational burden in unsupervised text clustering.

Main Methods:

• A contrastive learning module optimizes feature spaces by treating positive pairs and negative samples.
• Mixup data augmentation generates cost-effective virtual features, implicitly reducing computational load.
• The method involves selecting small data batches to simulate rare disease experimental conditions for effective clustering.

Main Results:

• The proposed method achieves superior experiment scores, particularly with small batch data, outperforming existing techniques.
• It effectively mitigates data overlap issues and enhances clustering accuracy for sparse medical text.
• Significant reductions in resource usage and time overhead were observed.

Conclusions:

• The contrastive clustering with mixup offers a favorable and effective strategy for unsupervised medical text clustering.
• This approach demonstrates cutting-edge outcomes, particularly beneficial for analyzing rare disease data.
• The method optimizes feature representation and enhances clustering performance efficiently.