graph construction

Existing heterogeneous graph neural networks (HGNNs) rely on predefined schemas and manual preprocessing for graphs lacking prior type information and exhibiting non-uniform feature formats, while large language model (LLM)-based approaches often neglect heterogeneity. Method: We propose LLM-GNN, a novel collaborative framework that enables end-to-end automatic format understanding, dynamic type induction, and cross-source feature alignment: an LLM performs semantic parsing of node/edge types to generate a structured schema; an adaptive module aligns heterogeneous features; and a lightweight GNN learns structured representations. Contribution/Results: Our method requires no type annotations or manual preprocessing. We provide theoretical guarantees on representation consistency and convergence. Evaluated on five standard heterogeneous graph benchmarks, LLM-GNN achieves an average 12.7% improvement in downstream task performance over state-of-the-art baselines.

This work addresses the challenges of feature heterogeneity and cross-domain transfer in graph data caused by the absence of textual information. It proposes “learnable graphlets” as the minimal semantic units of graphs, enabling for the first time a text-free cross-domain graph pre-training framework. By designing graphlet decomposition, a graphlet encoder, and an aggregator, the approach constructs a domain-agnostic architecture that extracts transferable knowledge from multi-domain graph data. The method supports joint pre-training across multiple domains and consistently achieves significant performance gains on diverse downstream tasks and datasets. Moreover, its effectiveness scales with the volume of pre-training data, and it reveals intrinsic connections between graphlet representations, existing graph models, and the transferability of node embeddings.

This work addresses the limitation of existing graph learning approaches, which typically operate in isolation within a single modality and task, thereby hindering the cross-task and cross-modal reuse of structural knowledge. To overcome this, the authors propose G-Substrate, a novel framework that models graph structures as persistent, shareable substrates. By unifying structural patterns and employing a role-interleaved training strategy, G-Substrate enables collaborative learning across multiple tasks and modalities. This approach facilitates the continuous accumulation and transfer of graph-structured knowledge, consistently outperforming both isolated training and conventional multi-task learning methods across diverse domains, modalities, and tasks.

Existing visual graph recognition methods are often confined to specific tasks and lack generalizability and cross-scenario transferability. This work proposes GraSP, an end-to-end framework based on subgraph prediction that jointly models graph structure and visual features to enable unified recognition of diverse graph types and rendering styles. GraSP achieves cross-task transfer without task-specific fine-tuning, representing the first general-purpose and transferable approach for visual graph recognition. Evaluated on multiple synthetic benchmarks and a real-world application, GraSP demonstrates exceptional generalization and adaptability, advancing the field toward a unified paradigm for graph recognition.

This work addresses the challenge of representation learning in heterogeneous graphs where heterophily and structural noise coexist. To tackle this issue, the authors propose HGUL, a unified framework that, for the first time, integrates heterophily modeling and structural denoising within a single end-to-end architecture. HGUL synergistically combines three components: kNN-based graph reconstruction, adaptive graph structure refinement, and heterogeneous affinity learning via polynomial graph kernels, effectively capturing cross-class node relationships while suppressing noise interference. Extensive experiments demonstrate that HGUL significantly outperforms existing methods across multiple benchmark datasets, exhibiting exceptional robustness in noisy scenarios. These results substantiate the efficacy and necessity of jointly modeling heterophily and structural noise for learning robust representations in heterogeneous graphs.

Unlike vision and language domains, graph learning lacks a shared input space, as input features differ across graph datasets not only in semantics, but also in value ranges and dimensionality. This misalignment prevents graph models from generalizing across datasets, limiting their use as foundation models. In this work, we propose ALL-IN, a simple and theoretically grounded method that enables transferability across datasets with different input features. Our approach projects node features into a shared random space and constructs representations via covariance-based statistics, thus eliminating dependence on the original feature space. We show that the computed node-covariance operators and the resulting node representations are invariant in distribution to permutations of the input features. We further demonstrate that the expected operator exhibits invariance to general orthogonal transformations of the input features. Empirically, ALL-IN achieves strong performance across diverse node- and graph-level tasks on unseen datasets with new input features, without requiring architecture changes or retraining. These results point to a promising direction for input-agnostic, transferable graph models.

This work addresses the excessive computational and memory overhead in existing graph structure learning methods, often caused by redundant edges. To mitigate this issue, the study introduces diversity into graph structure learning for the first time, proposing a novel edge construction strategy that jointly leverages node similarity and diversity. The resulting graph structure is optimized under mutual information guidance, enabling the method to function as a plug-and-play module compatible with prevailing frameworks. This approach significantly reduces the number of edges while simultaneously enhancing model performance. Extensive experiments demonstrate consistent and substantial performance gains across six state-of-the-art graph structure learning methods, validating the effectiveness and generalizability of the proposed technique.

Existing large language models struggle to effectively capture the high-order relational semantics inherent in hypergraphs due to their reliance on pairwise graph representations. To address this limitation, this work introduces a novel “hypergraph-as-language” paradigm and proposes Hyper-Align, a native hypergraph alignment framework that compiles hypergraph context into token sequences comprehensible to large language models through semantic-structural disentanglement and bidirectional message passing. Hyper-Align pioneers a linguistically grounded hypergraph representation by integrating a Hypergraph Incidence Detail Template (HIDT-O), a Hypergraph Incidence Projector (HIP), and a unified input protocol, enabling both vertex- and hyperedge-level question answering with frozen large language models. Evaluated on the newly curated HyperAlign-Bench benchmark, Hyper-Align significantly outperforms existing methods under both in-domain and zero-shot settings.

Existing knowledge graphs lack a unified structured vocabulary, which limits the transferability of foundation models across unseen graphs. This work proposes graphlets as universal structural tokens and introduces, for the first time, a systematic framework of model-agnostic structural vocabularies based on closed and open 2- and 3-paths as well as star-shaped graphlets. By leveraging pattern matching to uncover local topological invariances among relations, the approach enables zero-shot inductive and transductive link prediction. Evaluated across 51 cross-domain knowledge graphs, the method significantly outperforms current knowledge graph foundation models, demonstrating the effectiveness of graphlet-based vocabularies in enhancing cross-graph generalization.