heterogeneous gnn ranking

This paper identifies and quantifies a significant performance heterogeneity phenomenon in Graph Neural Networks (GNNs) for graph-level classification and regression: identical models exhibit substantial performance variance across individual graph samples, unexplained solely by topological differences. To address this, we propose a heterogeneity measurement framework based on Tree Mover’s Distance (TMD), the first to jointly model graph topology and node feature distributions. We further design a data-aware selective rewiring strategy and a spectral-adaptive depth selection mechanism. Experiments demonstrate that our approach improves average graph classification accuracy by 2.3% and substantially reduces hyperparameter tuning overhead. We empirically validate that performance heterogeneity strongly correlates with inter-class distance ratios. Moreover, our spectral-driven depth heuristic achieves performance on par with manually optimized layer counts across multiple benchmarks.

Existing heterogeneous graph neural networks (HGNNs) heavily rely on node/edge type labels for parameterization, leading to poor semantic generalizability, limited cross-type knowledge transfer, and weak interpretability. To address this, we propose the first integration of a Mixture-of-Experts (MoE) mechanism into the Heterogeneous Graph Transformer (HGT), introducing a type-agnostic, semantics-driven expert routing scheme. Specifically, we randomly mask type embeddings during training to attenuate reliance on superficial type labels, enabling experts to specialize according to intrinsic semantic patterns rather than predefined types. Evaluated on link prediction across IMDB, ACM, and DBLP, our method significantly outperforms standard HGT and type-aware MoE baselines. It achieves superior generalizability, higher computational efficiency, and enhanced interpretability—offering a novel lightweight, semantics-adaptive architectural paradigm for heterogeneous graph modeling.

Existing heterogeneous graph neural networks (HGNNs) suffer from parameter explosion and relation collapse—where the number of learnable parameters grows quadratically with the number of relation types, and discriminative capacity across relations degrades—hindering scalability to large-scale heterogeneous graphs with abundant relation schemas. To address this, we propose the Blend&Grind mechanism: a lightweight, scalable HGNN framework that models diverse relations efficiently using a single shared parameter set. It operates in a unified feature space via differentiable relation blending (Blend) and hierarchical relation refinement (Grind), augmented by relation-aware projection and structured regularization. Evaluated on multiple benchmarks, our method reduces model parameters to just 1/28.96 of the prior state-of-the-art, accelerates training throughput by 8.12×, and improves node classification accuracy by up to 7%, thereby substantially overcoming the scalability bottleneck of HGNNs.

This work addresses negative transfer in pre-trained heterogeneous graph neural networks (HGNNs) for semi-supervised node classification, caused by objective mismatch between pre-training and downstream tasks. We propose a novel “pre-training + prompting” paradigm. Our method introduces: (1) the first heterogeneous-graph-aware prompt function, jointly modeling virtual class prompts and heterogeneous feature prompts; and (2) a multi-view neighbor aggregation mechanism, enabling the first general adaptation of prompt tuning to HGNNs. Crucially, our approach requires no fine-tuning of model parameters—only lightweight prompt modules are optimized. Evaluated on three standard heterogeneous graph benchmarks, it consistently outperforms state-of-the-art HGNNs, achieving average accuracy gains of 3.2–5.8 percentage points. These results validate the effectiveness and generalizability of prompt-based alignment between pre-training objectives and downstream tasks.

Existing knowledge graph link prediction methods struggle to scale to *k*-ary relational modeling. Method: This paper introduces the first link prediction framework for relational hypergraphs. Contributions/Results: (1) We propose the first Graph Neural Network (GNN) architecture specifically designed for relational hypergraphs, and theoretically prove its expressive power is equivalent to the relational Weisfeiler–Leman algorithm and first-order logic. (2) The framework unifies inductive and transductive prediction, overcoming traditional limitations in modeling higher-order relations. Evaluated on multiple relational hypergraph benchmarks, our method significantly outperforms all baselines: it achieves substantial gains in inductive link prediction and attains state-of-the-art performance in transductive settings. This work establishes a new paradigm for higher-order relational learning—provably expressive, interpretable, and scalable.

This work addresses the heterogeneity inherent in real-world graph data by systematically investigating the impact of expert-level diversification strategies on Graph Neural Network (GNN) ensemble performance. We empirically evaluate 20 distinct diversification methods—including random reinitialization, architectural variation, directional modeling, data partitioning, and hyperparameter perturbation—across 14 node classification benchmarks, delivering the first comprehensive quantitative analysis of GNN expert diversity. Leveraging a Mixture-of-Experts framework, we construct and evaluate over 200 ensemble variants, uncovering a nonlinear relationship between diversity, complementarity, and generalization performance. We further propose a reproducible, principled training guideline for diversified GNN ensembles. Results demonstrate that judiciously introduced expert diversity significantly enhances both robustness and accuracy, with top-performing ensembles outperforming their best single-model baselines by up to 5.2% (average improvement). The implementation is publicly available.

Existing graph neural network pre-training methods primarily target homogeneous graphs and overlook semantic mismatch—a prevalent issue in heterogeneous graphs wherein raw data exhibits a semantic gap relative to ideal, transfer-rich representations. Method: We propose the first dual-aware pre-training framework for large-scale heterogeneous graphs. It jointly models heterogeneous topology via structure-aware pretext tasks and constructs semantic neighborhood perturbation subspaces through semantic-aware tasks, explicitly mitigating semantic mismatch in a self-supervised manner. The framework integrates heterogeneous structural modeling, semantic neighbor discovery, and perturbation subspace construction. Contribution/Results: Evaluated on multiple real-world million-scale heterogeneous graph datasets, our method consistently outperforms state-of-the-art approaches, yielding average improvements of 3.2–5.8 percentage points across downstream tasks. It significantly enhances model generalizability and cross-task knowledge transfer capability.

This work addresses the limitations of existing heterogeneous graph neural networks (HGNNs), which rely on a single shared linear decoding head and thus struggle to capture fine-grained semantic relationships, often overfitting to central nodes while neglecting long-tail ones. To overcome this, we propose HOPE—a plug-and-play heterogeneous graph decoding framework that dynamically routes nodes to semantically aligned expert decoders via a heterogeneity-aware prototype-based routing mechanism. HOPE further enforces orthogonality constraints among experts to enhance diversity and prevent expert collapse. By integrating prototype learning with a Mixture-of-Experts architecture, HOPE consistently boosts the performance of multiple state-of-the-art HGNN backbones across four real-world datasets, achieving significant gains with minimal computational overhead.

This work addresses the limitation of existing knowledge graph embedding methods, which primarily focus on binary relations and struggle to jointly model hyperedges and hyper-relations—two distinct types of n-ary facts commonly coexisting in real-world scenarios. To overcome this challenge, the authors propose HEHRGNN, a novel model that, for the first time, integrates both hyperedges and hyper-relations into a unified framework. The approach introduces a standardized HEHR representation format for n-ary facts and designs a tailored graph neural network message-passing mechanism capable of handling both structural types. This enables joint embedding learning and inductive link prediction over complex knowledge graphs. Extensive experiments on multiple real-world datasets demonstrate that HEHRGNN significantly outperforms current state-of-the-art baselines, confirming its superior capability in modeling heterogeneous n-ary facts and generalizing to unseen entities.

Graph Neural Networks (GNNs) often underperform on heterophilous graphs, where adjacent nodes exhibit significantly different labels or features. While existing approaches primarily focus on architectural modifications, they fail to fundamentally alleviate the underlying heterophily issue. This work proposes GRAPHITE, a novel framework that explicitly enhances graph homophily through direct graph transformation. Specifically, GRAPHITE introduces auxiliary feature nodes and reconstructs the graph structure based on a homophily-aware criterion, thereby optimizing message passing. Theoretical analysis and extensive experiments demonstrate that GRAPHITE substantially outperforms state-of-the-art GNN methods across multiple heterophilous graph benchmarks, while maintaining competitive performance on homophilous graphs.