Researchers propose unsupervised data augmentation method combining GMMs and LLMs to address imbalanced text clustering

Researchers from the University of Paris and University of Oran 1 propose a method to improve unsupervised text clustering in the presence of imbalanced data by combining Gaussian Mixture Models (GMMs) with Large Language Models (LLMs). The core idea is to use GMMs to identify clusters corresponding to underrepresented topics and then leverage LLMs to generate synthetic documents that enrich these clusters, thereby improving their representation in the dataset.

The authors argue that traditional clustering methods often fail to adequately capture minority topics in unsupervised settings, which can skew downstream analyses. Their approach aims to address this by first detecting minority clusters using the flexibility and robustness of GMMs, and then using LLMs to create synthetic documents that bolster the representation of these clusters.

In experiments across multiple imbalanced text datasets, the proposed method preserved clustering performance in all cases and often improved cluster interpretability. The authors describe the approach as robust and scalable, suggesting it could be broadly applicable to unsupervised NLP tasks where data imbalance is a concern.

The paper is titled 'Imbalanced Data Clustering via Targeted Data Augmentation Using GMM and LLM' and was submitted to arXiv on May 19, 2026. It is associated with the 23rd International Symposium on Intelligent Data Analysis (IDA 2025), where it is slated to appear in the proceedings.