Showing 13 papers for 2026-03-20
This work introduces RL-ASM, a Reinforcement Learning based approach to approximate subgraph matching. It learns neural graph representations and a policy to guide search for query subgraphs inside a large target graph, addressing NP-hardness and exploiting graph structure beyond heuristic methods. Experiments show improved accuracy and efficiency over traditional heuristic approaches.
This paper proposes Attack by Unlearning, a class of adversarial attacks that exploit approximate graph unlearning procedures for GNNs. As privacy regulations push for unlearning, the authors analyze how unlearning can degrade model performance and be exploited to cause privacy or security issues. They discuss potential defenses and practical guidelines.
This position paper argues that Spectral Graph Neural Networks are not truly spectral nor inherently superior for node classification. It identifies two theoretical flaws: the so-called graph Fourier bases are not classical Fourier bases, and high-degree polynomials can interpolate any spectral response, undermining key assumptions. The authors conclude that empirical success owes to factors beyond a faithful spectral interpretation.
This work proposes Training-Only Heterogeneous Image-Patch-Text Graph supervision for advancing few-shot learning adapters. Instead of modifying the lightweight adapter during inference, they train a high-capacity auxiliary Heterogeneous Graph Teacher to supervise fine-grained patch-text relationships. This training-only framework improves few-shot adapter performance while avoiding extra inference cost.
This paper presents WarPGNN, a parametric thermal warpage analysis framework powered by physics-aware Graph Neural Networks for SiP chiplet-based designs. GNNs provide fast, scalable predictions compared with full numerical simulations by modeling the thermal graph of the package. The approach achieves accurate warpage predictions with reduced computation, enabling parametric design exploration.
RewardFlow introduces a topology-aware reward propagation method on state graphs to facilitate RL with large language models. It addresses sparse terminal rewards by propagating rewards along the state graph guided by topology, enabling state-level credit assignment without heavy reward-model training. The approach is lightweight and improves agentic RL with LLMs.
CADGL presents Context-Aware Deep Graph Learning for predicting Drug-Drug Interactions. It addresses generalization under extreme cases, robust feature extraction, and real-world applicability by leveraging contextual graph representations and deep learning. The method aims to improve predictive performance and robustness over existing DDI models.
This paper discusses Sheaf Neural Networks (SNNs) and their biomedical applications. It presents the theory and mathematical modeling of SNNs and demonstrates their effectiveness on a biomedical case study, outperforming popular GNNs such as GCN, GAT, and GraphSage. The work argues that SNNs provide a principled framework for multi-scale and heterogeneous biomedical data.
This paper presents an innovative self-supervised heterogeneous graph neural network to address spatial resolution mismatches in energy system coupling. It models high-resolution geographic units as graph nodes and integrates multiple geospatial features to learn physically meaningful weights for aggregation. The approach improves spatial allocation accuracy and scalability of energy system models.
HypeMed proposes hypergraph-based modeling of patient relationships to enhance medication recommendations. It preserves visit-level combinatorial semantics while leveraging historical references through visit-conditioned retrieval. The method aims to generate safer and more effective medication sets from health records.
UGID introduces an internal-representation level debiasing framework by treating the Transformer as a structured graph and applying graph isomorphism concepts. It aims to remove biases embedded in representations, beyond data- or output-level debiasing. The framework shows improvements in debiasing while preserving model capabilities.
TDAD provides pre-change impact analysis for AI coding agents, building a dependency map between source code and tests. Before patching, agents know which tests to run, reducing regressions. It is open-source and aims to improve reliability of AI coding agents.
QuaQue designs and implements a condensed algebra for concurrent versioning of knowledge graphs, translating SPARQL queries into SQL. It represents versioning information compactly using bitstrings on a relational model, enabling efficient cross-version querying within the ConVer-G system.