Showing 21 papers for 2026-03-23
We introduce KD-Brain, a Prior-Informed Graph Learning framework that explicitly encodes domain priors to guide the learning of interactions among functional subnetworks in heterogeneous brain networks. By incorporating semantic-conditioned priors into a Transformer-based architecture, it improves generalization under limited training samples and yields more interpretable subnetwork interactions.
We propose MemReward, a graph-based experience memory for reward prediction when reward labels are scarce for LLM RLHF. The memory propagates limited human-provided rewards across an experience graph to improve reward estimation and downstream RL fine-tuning efficiency.
We study decision making on expanding graphs where nodes arrive with unknown patterns and the graph evolves over time. We develop a stochastic sequential decision framework that leverages graph filtering to account for evolution and uncertainty, yielding non-myopic decisions that consider future network impact.
GDEGAN introduces Gaussian Dynamic Equivariant Graph Attention for ligand binding site prediction on 3D protein structures. It modulates attention by capturing variations in chemical and geometric properties of neighbors, improving accuracy over standard dot-product attention in equivariant GNNs.
This work integrates dataset meta-features with knowledge graph embeddings to enhance meta-learning, focusing on pipeline performance estimation (PPE) and dataset performance-based similarity estimation (DPSE). The proposed method combines meta-features with KG embeddings to improve transfer and similarity judgments across datasets.
We propose a hypergraph neural network method for solving polynomial-objective integer programming, capturing high-degree variable interactions via a high-degree-term-aware hypergraph representation. This approach enables better modeling of nonlinear discrete decisions and shows improved solution quality on POIP problems.
We develop a graph neural network that predicts filtered production rates on non-uniform meshes for large-eddy simulations, using inputs of filtered mass fractions and temperature. The model is trained on DNS data of turbulent premixed hydrogen-methane jets with varying hydrogen fractions, using Favre filtering aligned to the mesh.
We present a structural controllability framework for hypergraphs to address large-scale systems with higher-order interactions. The paper derives conditions and algorithms for achieving controllability in hypergraphs beyond traditional exact controllability, enabling practical control design.
We propose a cellular sheaf-theoretic framework to analyze complex network models and attention mechanisms in GNNs and topological deep learning. The framework studies how signals distribute and diffuse during training, offering insights into aggregation, diffusion, and representation stability.
Hyper-STTN constructs multiscale hypergraphs to model group dynamics and higher-order interactions in crowds, and uses a Spatial-Temporal Transformer Network to predict trajectories. The method improves capture of heterogenous group influences and long-range dependencies for accurate trajectory forecasting.
AtGCN introduces a graph convolutional network for detecting ataxic gait from 2D video data and estimating its severity. The approach addresses subtle gait deviations and the small size of available datasets by leveraging graph-based spatial-temporal features.
We propose an unsupervised feature selection method combining a robust autoencoder with adaptive graph learning to simultaneously cluster data and identify discriminative features. The non-linear representations and robustness to outliers overcome limitations of linear projection and uniform cluster assumptions.
DIAL-KG enables schema-free incremental knowledge graph construction via dynamic schema induction and evolution-intent assessment. It supports dynamic data arrival without full recomputation and assesses the evolution intent to guide incremental updates.
We introduce dynamic belief graphs to support theory-of-mind reasoning in LLMs, modeling how people's beliefs evolve and influence actions under uncertainty. The approach yields coherent mental models over time for high-stakes contexts.
We propose GraphRAG, a retrieval-augmented generation approach that preserves structural relationships in educational content to improve automated short answer grading. By using a graph-based retrieval over multi-hop structures, it reduces hallucinations and better adheres to rubrics.
We present a graph neural network–driven co-design framework for jointly optimizing morphology and control in soft robots. By modeling body-brain co-evolution, the method mitigates disruption of learned control during morphology changes.
FinReflectKG-EvalBench provides a benchmark and evaluation framework for information extraction and KG construction from SEC 10-K filings. It emphasizes multi-dimensional evaluation principles to assess financial knowledge graphs.
CO-EVOLVE proposes bidirectional co-evolution between graph structure and semantics to improve heterophilous learning where neighboring nodes are dissimilar. It enables dynamic updating to reduce semantic-structural misalignment and enhances grounding.
We introduce graph-enhanced LLMs for e-commerce appeal adjudication, addressing information asymmetry by modeling verifiable reviewer actions. The framework imposes explicit action constraints to ground reasoning and improve decision quality.
RouterKGQA presents a specialized-general model routing framework for constraint-aware knowledge graph question answering, combining small, efficient specialized models with a large general model. The router selects the appropriate model to balance efficiency, grounding, and accuracy in knowledge-grounded QA.
The paper introduces a condensed representation for snapshot-based RDF graphs to efficiently capture evolving knowledge graphs updated periodically from heterogeneous sources. It formalizes the representation and addresses organizing evolving data to enable easy access and analysis, facilitating management of data evolution in knowledge graph systems.