Showing 21 papers for 2026-04-15
DBGL introduces Decay-Aware Bipartite Graph Learning to model irregular medical time series. By incorporating decay dynamics in a bipartite graph, it addresses irregular sampling, asynchronous observations, and varying gaps to produce more robust representations for classification.
Graph Concept Bottleneck (GCB) enables self-explainable text-attributed graph learning by mapping graphs into a concept bottleneck space where each concept is a meaningful phrase. Predictions are made from activations of these concepts, and the information bottleneck is used to prune irrelevant concepts.
XANE(3) is an E(3)-equivariant graph neural network for predicting XANES spectra directly from atomic structures. It combines tensor-product message passing with spherical-harmonic edge features, absorber-query attention pooling, and specialized normalization and residuals, trained with a composite loss to improve line-shape fidelity.
This survey reviews uncertainty quantification (UQ) methods for graph learning. It categorizes approaches, discusses how to quantify and calibrate uncertainty in graph models, and outlines open challenges and future directions.
ScaleNet introduces scale-aware message passing for node classification and formalizes scale invariance in graphs, providing theoretical guarantees and empirical results. Moving beyond fixed k-hop depths, it leverages multi-scale representations to improve performance.
We extend Cluster Catch Digraphs to moderate-dimensional data with Uniformity- and Neighbor-based CCDs. Our approach uses a nearest-neighbor-distance Monte Carlo spatial randomness test to adapt the covering radii, addressing limitations of Ripley’s K-based tests at higher dimensions.
HSG-12M presents a large-scale benchmark of spatial multigraphs derived from the energy spectra of non-Hermitian crystals. The dataset encodes Hamiltonian spectral graphs to enable AI-driven exploration of complex crystal physics.
We show that distributed ADMM iterations can be cast as message passing on graphs. A graph neural network is trained to accelerate convergence and automate hyperparameter tuning, yielding faster and more robust distributed optimization.
GTCN-G is a residual graph-temporal fusion network for imbalanced intrusion detection. It blends graph-based representations with temporal convolutions and uses gating to handle class imbalance, improving detection performance.
This work studies clean-label backdoor attacks on graph neural networks, where triggers are embedded without altering training labels. It analyzes how to poison the model’s inner prediction logic under clean-label constraints and demonstrates threat effectiveness.
We propose an efficient and scalable granular-ball graph coarsening method for large-scale node classifications. Granular-ball coarsening preserves multi-granularity information and reduces computation, enabling scalable GCN training on big graphs.
CoG proposes controllable graph reasoning over knowledge graphs using relational blueprints and failure-aware refinement. It is a training-free framework intended to improve reliability of KG-augmented LLMs under neighborhood noise and misalignment.
GAM introduces Hierarchical Graph-based Agentic Memory for LLMs to balance continuous information gathering and long-term knowledge retention. It uses a hierarchical graph memory to decouple transient context updates from durable knowledge.
KG-Reasoner proposes a reinforced model for end-to-end multi-hop knowledge graph reasoning in KBQA. By learning a reinforcement-guided reasoning policy, it avoids brittle pipeline decompositions.
Topology-Aware Reasoning over Incomplete Knowledge Graphs with Graph-Based Soft Prompting shifts from explicit edge traversal to subgraph-level reasoning via soft prompts. This makes multi-hop KGQA more robust to incomplete or noisy graphs.
Learning Chain Of Thoughts Prompts for Predicting Entities, Relations, and Literals on Knowledge Graphs introduces RALP, which learns string-based CoT prompts as scoring functions for KG triples. Bayesian optimization identifies effective prompts to guide reasoning.
Think Parallax reveals that multi-hop KG reasoning is naturally multi-view, as transformer heads specialize in distinct relations. It proposes retrieval-augmented generation that leverages multiple KG representations instead of collapsing hops into a single embedding.
HiFiNet is a hierarchical fault identification framework for wireless sensor networks that combines edge-based classification with graph aggregation. It leverages spatio-temporal correlations to improve accuracy while balancing energy consumption.
KG-Hopper empowers compact open LLMs with knowledge-graph reasoning via reinforcement learning. It enables effective multi-hop KGQA without requiring very large language models.
Decoupling Distance and Networks presents a hybrid spatio-temporal modeling approach that combines Graph Attention Networks with a latent Gaussian process. It jointly captures relational structure and spatial dependence for risk mapping and uncertainty quantification.
Session-based recommender systems often treat sessions as anonymous, which limits personalization in sparse or cold-start scenarios. This work proposes a persona-driven SBRS that explicitly models latent user personas inferred from interaction sequences. It leverages large language models to generate rich item representations and incorporates heterogeneous knowledge graphs to align recommendations with the inferred personas, aiming to improve personalization under data sparsity.