MASK: Minority-aware Semantic Alignment via Knowledge Enhancement for Long-tailed Graph Clustering

Abstract

Text-attributed graph (TAG) clustering seeks to identify node groups by jointly exploiting the semantic information encoded in textual attributes and the structural information captured by graph connectivity. Despite significant progress, their performance often becomes less satisfactory under long-tailed distributions. To address this challenge, we introduce Minority-Aware Semantic Alignment via Knowledge Enhancement (MASK), an unsupervised framework designed for long-tailed TAG clustering. The core idea of MASK is to integrate the textual understanding capability of large language models (LLMs) with canonical correlation maximization to construct minority-aware semantic groups. More specifically, MASK first leverages LLMs to augment textual attributes and generate multiple textual views, which are then aligned via canonical correlation. Subsequently, we randomly assign samples into balanced groups with mixed semantics for consistency learning and then emphasize those with higher semantic discrepancy as minority candidates. To enhance clustering with external knowledge, we leverage LLMs to fine-tune the model with ranking guidance, which is then formulated as a regularization term. Extensive experiments on four datasets under two varying imbalance ratios demonstrate the effectiveness of MASK. The implementation of MASK is publicly available at https://github.com/fuyw-aisw/MASK.