Overcoming Data Sparsity in Social Computing: A GMM+LLM Approach

Abstract

The rise of social media as the primary platform for high school students’ interactions presents both opportunities and challenges for social computing research. While social media offers rich insights into adolescent social behavior, research often suffers from insufficient sample sizes, leading to biased portrayals of social dynamics. To address this limitation, this study proposes a novel framework that integrates Gaussian Mixture Models (GMM) with Large Language Models (LLMs) to generate linguistically detailed and context-sensitive synthetic data. Using a dataset of 15,000 student social profiles collected from 2006 to 2009, GMM clustering was applied to identify distinct user groups, followed by LLM-driven content generation tailored to each cluster through structured prompt engineering. Comparative analysis with K-means clustering demonstrates GMM’ s superior ability to handle outlier-prone data, producing smoother and more representative clusters. The synthesized content successfully reflects authentic patterns of adolescent engagement across domains such as sports, music, religion, and negative topics, thereby mitigating the problem of data sparsity. This study highlights the potential of combining probabilistic modeling with advanced natural language generation to enhance the authenticity and scalability of social dynamics simulation. Future research should expand contextual features and employ robust evaluation mechanisms to further refine generative accuracy.