Can Representation Gaps Be the Key to Enhancing Robustness in Graph-Text Alignment?

TL;DR

This work investigates why forcing graph-text alignment can degrade performance in TAGs and demonstrates that representation gaps act as protective geometry for modality-specific knowledge. It introduces LLM4GTA, a gap-aware alignment framework that preserves these gaps with an adaptive gap preservation module and enhances discriminative power via intra-graph-space compensation. Through theoretical analysis and extensive experiments, the method shows significant gains in zero-shot and few-shot transfer across diverse domains, validating that maintaining moderate gaps preserves both topological inductive biases and semantic priors. The approach offers a practical, data-efficient pathway for robust cross-modal learning in text-rich graphs, with a dual-space prediction fusion mechanism that balances cross-modal semantics and graph-space discrimination.

Abstract

Representation learning on text-attributed graphs (TAGs) integrates structural connectivity with rich textual semantics, enabling applications in diverse domains. Current methods largely rely on contrastive learning to maximize cross-modal similarity, assuming tighter coupling between graph and text representations improves transfer performance. However, our empirical analysis reveals that both natural gap expansion and forced gap reduction result in performance degradation by disrupting pre-trained knowledge structures and impairing generalization. This arises from the geometric incompatibility between encoders, where graph encoders capture topological patterns, while text encoders capture semantic structures. Over-alignment compresses these distinct spaces into shared subspaces, causing structure collapse that diminishes both topological reasoning and semantic understanding. We propose \textbf{LLM4GTA}, a gap-aware alignment framework that preserves representation gaps as geometric necessities for maintaining modality-specific knowledge and improving transfer performance. LLM4GTA includes an adaptive gap preservation module to prevent over-alignment by monitoring similarity evolution and an intra-modal compensation mechanism that boosts discriminative power using auxiliary classifiers in graph space. Extensive experiments show significant improvements over existing methods in zero-shot and few-shot scenarios.

Figures (6)

• Figure 1: Representation similarity and accuracy evolution on PubMed. Top: Naive task-specific fine-tuning causes representation drift, with both similarity and accuracy declining. Bottom: Contrastive alignment loss increases graph-text similarity but struggle to prevent accuracy degradation.
• Figure 2: Comparison of graph-text alignment methods for text-attributed graphs. (a) LLaGA employs a trainable projector to map GNN-encoded graph tokens into LLM embedding space. (b) TEA-GLM performs token-level alignment between GNN representations and LLM embeddings. (c) Our method builds upon self-supervised contrastive pretraining with LLM-generated summaries, introducing gap-aware adaptation through modality gap monitoring for training stability and dual-space classification for enhanced task adaptability.
• Figure 3: Impact of number of shots on node classification on three target domains.
• Figure 4: LLM4GTA training curves showing stable gap preservation with consistent accuracy gains, validating our gap-aware alignment approach.
• Figure 5: Generalization performance of LLM4GTA on zero-shot node classification across multi-dataset transfer settings.