Heterogeneous Graph Alignment for Joint Reasoning and Interpretability

Multi-graph learning is crucial for extracting meaningful signals from collections of heterogeneous graphs. However, effectively integrating information across graphs with differing topologies, scales, and semantics, often in the absence of shared node identities, remains a significant challenge. We present the Multi-Graph Meta-Transformer (MGMT), a unified, scalable, and interpretable framework for cross-graph learning. MGMT first applies Graph Transformer encoders to each graph, mapping structure and attributes into a shared latent space. It then selects task-relevant supernodes via attention and builds a meta-graph that connects functionally aligned supernodes across graphs using similarity in the latent space. Additional Graph Transformer layers on this meta-graph enable joint reasoning over intra- and inter-graph structure. The meta-graph provides built-in interpretability: supernodes and superedges highlight influential substructures and cross-graph alignments. Evaluating MGMT on both synthetic datasets and real-world neuroscience applications, we show that MGMT consistently outperforms existing state-of-the-art models in graph-level prediction tasks while offering interpretable representations that facilitate scientific discoveries. Our work establishes MGMT as a unified framework for structured multi-graph learning, advancing representation techniques in domains where graph-based data plays a central role.

💡 Research Summary

The paper introduces the Multi‑Graph Meta‑Transformer (MGMT), a unified framework for learning from collections of heterogeneous graphs that may differ in topology, scale, and node identity. MGMT first encodes each individual graph with a depth‑aware Graph Transformer (GT) that stacks multiple attention layers and aggregates their outputs using learned confidence scores, thereby capturing information from various receptive fields. From the attention maps produced by the GT, a small set of task‑relevant “supernodes” is selected per graph based on a threshold τ. These supernodes form pruned subgraphs that retain intra‑graph connectivity. All supernodes across the collection are then assembled into a meta‑graph. The meta‑graph contains two types of edges: (1) intra‑graph edges inherited from the original subgraphs, preserving local structure, and (2) inter‑graph edges added between supernodes from different graphs whose latent embeddings have cosine similarity above a second threshold γ, effectively aligning functionally similar substructures. Additional GT layers are applied to the meta‑graph, enabling fine‑grained message passing both within and across graphs. The authors provide theoretical results showing that the depth‑aware aggregation can recover L‑hop neighborhood mixing and that the function class induced by the meta‑graph strictly expands the approximation capacity compared with late‑fusion methods that only combine pooled graph embeddings. Empirical evaluation on synthetic benchmarks and real‑world neuroscience data (multi‑subject brain connectivity for memory and Alzheimer’s disease prediction) demonstrates that MGMT consistently outperforms state‑of‑the‑art multi‑graph and transformer models in classification accuracy and AUC. Moreover, the meta‑graph offers built‑in interpretability: the identified supernodes correspond to biologically meaningful brain regions, and the superedges reveal cross‑subject alignments that match known neural circuits. The paper concludes that MGMT is a scalable, backbone‑agnostic, and interpretable solution for structured multi‑graph learning, opening avenues for further research on dynamic meta‑graph construction and broader multimodal integration.