Title: Optimizing Text-Attributed Graph Condensation Through Strategic Text Selection and Attribute Alignment

Text-Attributed Graphs (TAGs), a significant category of structured data where every node is accompanied by a textual description, present unique challenges. Typically, models handling TAGs employ joint training of Graph Neural Networks (GNNs) and language models, a process that demands substantial computational resources in terms of both time and memory, particularly when processing extensive datasets. To address these efficiency bottlenecks, we introduce TAGSAM, a novel condensation technique designed to shrink TAGs without compromising training efficacy.

TAGSAM incorporates two primary mechanisms: subgraph text selection and attribute similarity matching, which target the compression of textual data and graph topology, respectively. In the domain of text, the subgraph text selection process identifies and consolidates key text segments from various related descriptions. This is achieved by maximizing mutual information to ensure the retention of representative content. Regarding graph topology, existing condensation approaches relying on Matching Training Trajectories (MTT) often struggle with high variance, which negatively impacts accuracy. TAGSAM resolves this by aligning stable similarity matrices through its attribute similarity matching component.

In comparative evaluations against six leading baseline methods, TAGSAM demonstrated superior results. When constrained to the same compressed size, it outperformed the top-performing baseline by an average margin of 4.9% in accuracy. Moreover, the method proves robust enough to maintain competitive training performance even when the TAG is reduced to merely 1% of its original size. The source code for TAGSAM is publicly accessible at https://github.com/SundayVHan/TAGSAM.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC