New Framework Boosts Scalability for Text-Attributed Graph Learning

Text-Attributed Graphs (TAGs), which integrate graph topology with rich textual semantics, are powerful data models, but existing representation learning methods face severe scalability issues, particularly when combined with Large Language Models (LLMs). Current data distillation techniques also fall short in capturing the intricate interplay between graph and text modalities, struggle with limited labels in semi-supervised settings, and cannot produce human-readable textual attributes needed for LLM-based tasks. To overcome these challenges, a unified semi-supervised framework called \algo{} has been proposed. Grounded in empirical findings, \algo{} features a graph-text collaborative encoding module that uses dual-pathway encoders (graph-aware and graph-free) within a self-training scheme to generate reliable pseudo-labels and fuse complementary graph-text features. Additionally, it includes a theoretically sound Wasserstein Distance-based graph sketching algorithm and a cost-effective LLM text synthesis module that leverages cluster-based keyword extraction to create coherent, human-readable summaries for condensed nodes. Extensive experiments show \algo{} achieves state-of-the-art performance-compression trade-offs for both GNN- and LLM-based downstream tasks, enabling efficient and effective TAG learning and analytics.