AGE: Adaptive-Masking for Graph Embedding in GraphRAG

AGE addresses the latent-feature mismatch between graph representations and frozen language models in GraphRAG. It uses a Transformer-based masked self-supervised learning framework designed in the style of text embedding encoders. Because graph key nodes may contain dominant contextual information and are difficult or inefficient to reconstruct from local context, AGE introduces a learnable node sampler that focuses masking and prediction on non-key nodes. The authors report improved accuracy for GraphQA systems using non-parametric search components across four benchmark datasets with different characteristics.