Showing 20 papers for 2026-04-22
FASE proposes a fairness-aware spatiotemporal event-graph framework for predictive policing. It builds a graph of Baltimore's 25 ZIP Code Tabulation Areas using about 140k Part 1 crime incidents at hourly resolution to forecast risk. The system jointly optimizes fairness-constrained patrol allocation and includes a closed-loop deployment feedback simulator to study bias and policy impact.
We propose a structure-aware multi-level temporal graph network (MTGN) with local-global fusion for industrial fault diagnosis. The model captures local, global, and dynamic sensor relationships in large-scale systems, addressing non-Euclidean graph structures. It improves fault detection and diagnostic accuracy by leveraging multi-level temporal correlations.
We introduce Subgraph Concept Network, which brings concept-level explanations to graph classification. Unlike prior approaches that explain node embeddings, it uses subgraphs to reveal higher-level concepts, and it preserves interpretability through less information loss caused by pooling.
Learning Posterior Predictive Distributions for Node Classification from Synthetic Graph Priors presents a Bayesian-style framework to obtain posterior predictive distributions for node labels using synthetic graph priors. This enables better generalization across graphs with varying properties, reducing the need for graph-specific labeled data. The approach improves universal node classification across diverse graphs.
Inductive Subgraphs as Shortcuts shows that recurring inductive subgraphs act as spurious shortcuts that mislead GNNs on heterophilic graphs. The work provides empirical and theoretical evidence of how these subgraphs reinforce misleading patterns, and it proposes causal disentanglement strategies to improve robustness.
The Logical Expressiveness of Topological Neural Networks investigates the expressiveness of TNNs. It argues that incorporating higher-order relational structures into message passing yields greater representational power than traditional GNNs, with formal insights into the logical capabilities of TNNs.
Revisiting Catastrophic Forgetting in Continual Knowledge Graph Embedding reassesses forgetting in CKGE. It shows that when new entities are added, their embeddings can interfere with previously learned tasks, not only erasing old knowledge. The paper provides theoretical and empirical analysis and suggests directions to mitigate interference.
This study critically re-evaluates GNNs for Bitcoin fraud detection under temporal distribution shift with leakage-free evaluation. In a strictly inductive setting, a feature-based Random Forest can outperform evaluated GNNs, challenging the presumed superiority of graph-based methods.
Towards Generalization of Graph Neural Networks for ACOPF proposes HH-MPNN, a Hybrid Heterogeneous MPNN that combines heterogeneous GNNs with a scalable transformer and physics-informed positional encodings. The architecture aims to generalize across different grid topologies while achieving faster and more reliable ACOPF solutions.
Graph Data Augmentation with Contrastive Learning on Covariate Distribution Shift studies covariate shift in graphs and introduces augmentation techniques plus contrastive learning to bolster GNN robustness under distribution shifts. The approach leverages latent-space information to better align training and test distributions.
Opinion de-polarization in social networks with GNNs investigates methods to reduce polarization in networks with echo chambers. The paper explores strategies to increase cross-community connectivity and information flow to mitigate polarization and echo-chamber effects.
VoteGCL introduces a data-augmentation framework for graph-based recommendations that uses LLMs to generate and rerank candidate items, and combines results via majority voting. This yields high-confidence synthetic user-item interactions and improves recommendation performance under data sparsity and popularity bias.
Temp-R1 is an autonomous end-to-end agent for temporal KGQA trained with reverse curriculum reinforcement learning. To cope with long reasoning, it expands the action space with specialized internal actions, enabling more scalable complex temporal reasoning.
DW-Bench provides a benchmark for LLMs on graph-topology reasoning over data warehouse schemas, including foreign-key and data-lineage edges. Experiments show tool-augmented LLMs outperform static baselines but still struggle on hard compositional subtypes.
Impact of large language models on peer review opinions analyzes how LLMs influence evaluative aspects of AI conference peer reviews, including linguistic form and scoring. The findings indicate that LLMs affect both assessment judgments and writing styles, with implications for academic integrity and quality.
Reasoning Over Space (ROS) proposes geographic-aware reasoning for LLM-based next-POI recommendation. It introduces a Hierarchical Spatial Semantic ID that discretizes locality and POI semantics into tokens, enabling the LLM to reason about space.
Modular Representation Compression studies how to compress LLM representations for efficient recommendations. It reports a counterintuitive phenomenon in mid-layer representations and proposes modular compression methods to balance efficiency and performance.
Spatiotemporal Link Formation Prediction in Social Learning Networks uses graph neural networks to predict future student interactions in educational settings. The work designs a GNN-based link predictor tailored to SLNs to support timely instructional interventions.
STK-Adapter introduces a method to incorporate evolving graphs and event chains for temporal knowledge graph extrapolation. It preserves spatial-temporal information during alignment with LLMs and mitigates information loss via an adapter mechanism.
SAGER presents Self-Evolving User Policy Skills for a recommendation agent. It addresses the static prompting bottleneck by enabling per-user memory and policy refinement, letting the agent evolve its reasoning skills as user interactions accumulate.