Large Language Models based Graph Convolution for Text-Attributed Networks
Keywords: Text attributed graphs, Long-context model
TL;DR: we propose a novel framework to adapt the long-context model for graph learning by retrieving both structural and text-related content.
Abstract: Text-attributed graph (TAG) tasks involve analyzing both structural information and textual attributes. Existing methods employ text embeddings as node features, and leverage structural information by employing Graph Neural Networks (GNNs) to aggregate features from neighbors. These approaches demand substantial computational resources and rely on two cascaded stages, limiting scalability in large-scale scenarios and making them vulnerable to the influence of irrelevant neighboring nodes. The advancement of language models (LMs) presents new avenues for tackling this task without GNNs, leveraging their ability to process text attributes of both the target node and its important neighbors. Instead of using graph convolution modules, LMs can assign weights to these tokens based on relevance, enabling token-level weighted summarization. However, it is nontrivial to directly employ LMs for TAG tasks because assessing the importance of neighbor nodes involves both semantic and structural considerations. Additionally, the large search space presents efficiency issues for computing importance scores in a scalable manner.
To this end, we propose a novel semantic knowledge and Structural Enrichment framework, namely SKETCH, to adapt LMs for TAG tasks by retrieving both structural and text-related content. Specifically, we propose a retrieval model that identifies neighboring nodes exhibiting similarity to the target node across two dimensions: structural similarity and text similarity. To enable efficient retrieval, we introduce a hash-based common neighbor estimation algorithm for structural similarity and a nearest-neighbor recalling algorithm for embedding similarity. These two similarity measures are then aggregated using a weighted rank aggregation mechanism. The text attributes of both the retrieved nodes and the target node provide effective descriptions of the target node and are used as input for the LM predictor. Extensive experiments demonstrate that SKETCH can outperform other baselines on three datasets with fewer resources.
Primary Area: other topics in machine learning (i.e., none of the above)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3664
Loading