Showing 7 papers for 2026-05-02
OptimusKG introduces a multimodal biomedical labeled property graph built from structured and semi-structured resources to preserve factual, type-specific metadata across molecular, anatomical, clinical, and environmental domains. It unifies disparate biomedical resources into a unified representation, containing 190,531 nodes, to maintain schema-level constraints and enable cross-domain reasoning.
This work tackles robust representation learning on heterogeneous, heterophilous graphs when connectivity is noisy or misleading. It identifies structural noise as a key degradation factor and proposes a graph-structure learning approach that learns to prune or reweight edges to improve downstream performance across diverse node types.
This paper presents a framework that constructs knowledge graphs from AI policy documents to support policy-compliance reasoning with LLMs. By building KGs from three AI-risk policies under two ontology schemas and evaluating five LLMs on 42 policy QA tasks across six reasoning types, the work demonstrates the value of structured policy knowledge for LLM reasoning.
This work investigates using large language models to refine the graph structure derived from EEG signals for seizure diagnosis. By leveraging LLM reasoning and contextual understanding, it aims to suppress irrelevant connections and highlight clinically meaningful relationships, improving representation learning and downstream detection accuracy.
This paper argues that tying all facts to a single decay rate is suboptimal since knowledge types exhibit distinct temporal dynamics. It proposes a hierarchical, continuous decay surface parameterized by two orthogonal signals—velocity and recency—to adaptively decay facts in a knowledge graph for retrieval.
This paper introduces TypeBandit, a lightweight, model-agnostic method for completing missing attributes in heterogeneous graph neural networks. It formalizes type-level information asymmetry and uses topology-aware initialization and type-wise reweighting to allocate context by node type, improving attribute completion across diverse types.
This paper tackles scalable traffic forecasting on large road networks by reducing the computational burden of graph convolutions. It proposes a regularized adaptive graph convolution framework that preserves spatial-temporal dependencies while lowering complexity, enabling accurate and efficient predictions on large-scale networks.