Showing 16 papers for 2026-07-10
We propose a graph-regularized deep learning approach for EEG-based emotion recognition. Emotions are modeled as graph nodes with edges encoding psychological proximity based on dimensional theories. The method introduces three complementary regularization strategies, including Graph Label Smoothing and a graph-based commuting distance regularizer, to exploit inter-emotion relationships and improve recognition.
We present path_boost, a Python package for interpretable graph-level prediction. PathBoost learns an additive model over path-based features by greedily discovering predictive labeled paths during training. To avoid enumerating all paths, the algorithm efficiently searches for informative paths, yielding interpretable substructure-driven predictions.
EdgeRefine studies privacy-utility trade-offs under edge differential privacy. It uses Jaccard sampling to perturb graph structure, balancing privacy with utility by reducing noise while preserving predictive performance for graph learning tasks.
We examine UMAP's internal kNN graph and show its value for sensemaking beyond the 2D embedding. Applying standard graph algorithms to the original kNN graph—before distortion—can identify representative data points, reveal structure, and support tasks like clustering and ranking.
We propose Context Graphs for proactive enterprise agents. The Context Graph is a live relational data structure modeling enterprise entities, their relationships, and temporal state transitions. Built atop this graph, Delta Dynamics enable proactive surfacing of relevant, actionable information before questions arise to boost productivity.
We introduce a hippocampus-inspired multi-view hypergraph learning framework for web finance fraud detection. By combining cognitive neuroscience inspiration with multi-view hypergraphs, the approach captures diverse relational views and rare fraud patterns, improving detection compared with standard GNNs.
KG-SoftMAP integrates soft knowledge-graph priors into Bayesian network structure learning from sparse discrete data. It encodes a knowledge graph as a finite-strength prior and adds this logit-form prior to the BDeu score, enabling MAP structure learning with informative priors and better recovery under sparsity.
We build a knowledge graph to portrait scientific and technological academic conferences, integrating papers, authors, institutions, projects, and topics. The graph enables efficient information extraction, retrieval, and analytics to construct accurate portraits of conferences and participants.
MetaHGNIE introduces meta-path induced hypergraph contrastive learning for heterogeneous knowledge graphs. By leveraging meta-relational structures to form higher-order hypergraphs, it captures richer interactions and decouples topology from semantic attributes to improve node importance estimation.
InductWave develops inductive multi-hop logical query answering for knowledge graphs. It extends beyond existential first-order queries to handle conjunction, disjunction, and negation, enabling reasoning over unseen entities in inductive settings with limited training data.
STAGformer is a Spatio-temporal Agent Graph Transformer for micro-mobility demand forecasting. It achieves linear computational complexity with a two-step agent attention mechanism, where learnable tokens first aggregate global information and then broadcast it, enabling efficient global modeling on large urban networks.
Hypergraph Neural Stochastic Diffusion presents an SDE framework for uncertainty estimation in hypergraph neural networks. It models uncertainty arising from attributes, labels, and incidence structures, using stochastic diffusion on hypergraphs to quantify predictive uncertainty beyond point estimates.
When Prompts Ignore Structure proposes ARGTCA, a graph-based attribute reasoning approach for calibrated VLMs. It builds a Symbolic Attribute Graph with class-attribute pairs as nodes and trains a graph attention network to reason over relational attributes for better calibrated predictions and reduced overconfidence.
Stability of Flow Models for Graph Signals analyzes permutation-equivariant flow-based generators on graphs. It shows that permutation equivariance is preserved in continuous normalizing flows parameterized by GNNs and discusses how structural perturbations influence stability.
Danus orchestrates mathematical reasoning agents with a shared fact-graph memory. It provides an orchestration system that coordinates parallel proof search and organizes intermediate claims via a global fact graph to improve reliability and scalability of research-level reasoning.
MetaConfigurator: AI-Assisted RDF Authoring from JSON Data describes an RDF authoring view that extends a JSON Schema editor to convert JSON/YAML/CSV into RDF using AI-assisted RML mappings. It supports refining triples and executing SPARQL queries to promote semantic interoperability.