Showing 16 papers for 2026-04-30
This paper surveys multi-agent deep reinforcement learning approaches that rely on graph neural networks to enable inter-agent communication along interaction graphs. It discusses how GNN-based communication can enrich agents' representations and improve coordination. It also notes a lack of a unified framework to classify and distinguish MARL methods with graph-based communication.
The authors show that, for a fixed graph, GNNs trained for link prediction do not generally learn representations that align with those learned for node classification. Instead, many models exploit a mini-batch dependent heuristic introduced by batch normalization to solve edge classification, which can hinder transferability.
We introduce MomentumGNN, a graph neural network architecture designed to conserve momentum by construction, improving the accuracy of predicting the temporal evolution of linear and angular momentum in deformable-object dynamics. Unlike prior GNNs that output unconstrained accelerations, MomentumGNN enforces momentum constraints throughout the computation.
The paper proposes an unsupervised graph-based framework for anomaly detection in accounting subject relationships. It models accounting subjects as graph nodes and encodes co-occurrence and debit/credit relationships as weighted edges to discover stable subject correspondences and detect structural deviations in ledgers and vouchers.
STLGT encodes traces as span graphs and uses a structure-aware linear graph transformer to forecast tail latency (p95) across multi-step service calls. It aims to capture long-range dependencies and non-stationary workloads with scalable inference for per-API tail-latency prediction.
PiGGO introduces Physics-Guided Learnable Graph Kalman Filters for virtual sensing of nonlinear dynamic structures under uncertainty. It combines physics-informed modeling, graph neural networks, and Bayesian state estimation to improve online state estimation under sparse sensing.
We study sparse graph learning from sparse data by maximizing the Fiedler number as a robust regularizer to encourage connectivity. A greedy algorithm iteratively adds/removes a single edge to produce a sparse, connected graph from undersampled observations.
To scale STGCN to large road networks, the paper proposes a regularized adaptive graph convolution that reduces quadratic complexity while maintaining accurate spatial-temporal traffic forecasting. This approach enables efficient inference on large-scale networks.
We propose a deep graph convolutional network framework for crime hotspot prediction that captures complex spatial dependencies among crime events. By modeling events on a graph, it addresses limitations of KDE and SVM approaches that treat spatial data as independent.
ARK is an adaptive knowledge-graph retriever that gives a language model control over the breadth–depth tradeoff in retrieval. It enables efficient, multi-hop evidence gathering by balancing wide coverage with focused traversal.
PACIFIER reframes opinion polarization moderation as a graph-based sequential planning task under Friedkin-Johnsen dynamics. It proposes a unified graph-learning framework to pace interventions for reducing polarization.
The paper presents a probabilistic graphical model built with graph neural networks for Bayesian inversion of discrete structural component states from measurements. It tackles ill-posedness by learning the likelihood and posterior in a data-driven, graph-based framework where the discrete states map to observed responses.
Deterministic Legal Agents proposes a canonical primitive API to enable auditable reasoning over temporal knowledge graphs for legal reasoning with LLMs. It provides structured interaction with structure-aware representations to preserve provenance and controllability.
PBiLoss introduces a popularity-aware regularization to mitigate popularity bias in graph-based recommender systems. It improves fairness and diversity by balancing exposure across items.
Graph Propagated Projection Unlearning (GPPU) is a unified, scalable method for class-level unlearning that works across vision and audio models. It uses graph-based propagation to identify class-specific directions in feature space and projects representations onto an orthogonal subspace, followed by targeted fine-tuning.
CacheRAG presents a semantic caching system to augment Retrieval-Augmented Generation for Knowledge Graph Question Answering. It caches retrieval plans and prior evidence to improve retrieval consistency and reduce schema hallucinations, enabling more robust KGQA.