GraphVec: Cross-Domain Graph Vectorization for Graph-Level Representation Learning

Learning transferable graph representations across heterogeneous domains is challenging because graph datasets differ in topology, node-attribute semantics, feature dimensions, and attribute availability. We propose GraphVec, a language-model-free framework that maps heterogeneous graphs into fixed-dimensional representations for graph-level tasks. GraphVec constructs multi-scale global relational graphs and spectral embeddings to replace incomparable raw attributes with comparable relational features, and further aligns them through a density-maximization orthogonal transformation with provable convergence. A GIN–Graph Transformer backbone equipped with hierarchical reference distribution encoding captures node-distribution information beyond standard pooling. We further establish an end-to-end generalization bound connecting cross-domain contrastive pretraining to few-shot classification on unseen domains. Experiments on 13 datasets against more than 15 baselines show that GraphVec consistently performs strongly in cross-domain few-shot graph classification and graph clustering, while also yielding competitive representations for few-shot node classification.