LLM Features Can Hurt GNNs: Concatenation Interference on Homophilous Graph Benchmarks
Research published on arXiv indicates that incorporating features generated by large language models (LLMs) into graph neural networks (GNNs) can, counterintuitively, decrease performance on certain homophilous graph benchmarks. This degradation occurs specifically when LLM features are added through simple input concatenation rather than more integrated training methods. The study proposes a metric called Delta_sig to predict when this concatenation interference is likely to occur, finding it correlates more strongly with performance drops than graph homophily. AI
IMPACT This research suggests that simply concatenating LLM-generated features may not always improve GNN performance, highlighting the need for careful integration strategies.