Showing 25 papers for 2026-05-28
This work moves DDI prediction from binary detection to predicting interaction mechanism types. It systematically compares three GNN architectures on a large benchmark with 38,337 positive drug pairs and 86 interaction types under identical training conditions. The study highlights cross-attention and siamese dual message passing and includes an acetylsalicylic acid validation to demonstrate practical mechanism prediction.
We study label-free learning on graphs when labels are provided by LLM annotators, which introduce noise. LLM errors are class-dependent and depend on graph context; this leads to systematic mislabels. We propose robust learning methods that model annotation noise and exploit graph structure to improve node classification without labels.
We propose temporal hyperbolic graph representation learning to model scale-free Internet routing graphs. Hyperbolic geometry captures hierarchical structure; the model handles evolving topology and heavy tailed RTT distributions. Experiments show improved RTT prediction and routing QoS over Euclidean TGNN baselines.
We address limitations of self-supervised graph clustering by Robust Contrastive Graph Clustering with Adaptive Local-Global Integration. The method flexibly captures high-order local structures and global semantics by adaptively integrating local and global signals. It outperforms state-of-the-art clustering methods on real-world graphs.
We introduce dynamic topic modeling with a higher-order hypergraphical representation. Traditional topic models use a single topic vector, limiting word dependencies; our higher-order hypergraph captures richer word interactions over time. This yields improved dynamic topic tracking and semantics.
We propose the generalized Tikhonov layer for interpretable-by-design GNNs. The propagation is R = (p(L) + Q)^{-1} Q where L is the normalized Laplacian, Q is a diagonal node importance, and p is a learnable polynomial. The learned parameters directly reveal which features and which parts of the graph the model used for predictions.
We explore applications of temporal graph learning to predict dynamics of biological systems. Current single-cell foundation models are static, lacking explicit temporal evolution of developmental programs. This work-in-progress investigates how temporal graphs can model how cellular states emerge and differentiate over time.
We present BIRDNet, which mines Boolean implication relationships from data into a typed directed knowledge graph and encodes it as a layered neural network where each hidden unit represents a mined rule linking two features. This design yields interpretable neural computations with rule-based behavior.
We revisit graph autoencoders as implicit contrastive learners. Structure-based and feature-based GAEs can be seen as implicit graph contrastive learning, differing mainly in how views are constructed. This perspective unifies GAEs and informs better design of self-supervised graph representations.
We provide a sheaf-theoretic and topological perspective on complex network modeling and attention in graph neural models. The cellular sheaf framework helps analyze how signals distribute and diffuse during training, linking local feature aggregation to global attention, and offering theoretical insights for geometric and topological deep learning and attention mechanisms.
GFMate enables test-time prompt tuning for Graph Foundation Models, addressing the problem that domain information is often entangled with pre-training. By tuning prompts at test time, GFMs can adapt to new domains with few-shot prompts, improving performance without retraining.
We introduce B-cos GNNs, a class of inherently explainable GNNs whose predictions decompose into per-node, per-feature contributions via a single input-dependent linear map. They replace nonlinear message and update functions with B-cos transforms, enabling instance-level explanations with one forward and backward pass.
GOProteinGNN leverages protein knowledge graphs to improve protein representations by integrating knowledge about proteins beyond amino acid sequences. The approach uses unsupervised representation learning with KG information to enhance function and interaction predictions.
This paper reviews GNN-based source detection methods and provides a benchmark study to improve methodological clarity and reproducibility. It surveys epidemic source localization on contact networks and advocates standardized tasks, datasets, and evaluation protocols.
HGMEM proposes hypergraph-based working memory to improve multi-step RAG for long-context relational modeling. The memory captures high-order correlations among primitive facts, enabling coherent multi-step reasoning and reducing fragmentation in RAG systems.
FundaPod introduces a multi-persona agent pod platform with a knowledge graph memory for AI-assisted fundamental investment research. The platform helps analysts gather evidence, compare viewpoints, and generate transparent, reusable investment memos.
Let Relations Speak presents an end-to-end LLM-GNN soft prompt framework for fraud detection that avoids hard graph textualization and handles multi-relational data. The soft prompts guide GNN reasoning to improve detection performance when textual data is limited.
LegalGraphRAG proposes multi-agent graph retrieval augmented generation for reliable legal reasoning, addressing heterogeneity and granularity in legal corpora by distinguishing facts, applied rules, and abstract principles, and using graph-based retrieval to support rigorous reasoning.
MetaboT introduces an LLM-based multi-agent framework for interactive analysis of mass spectrometry metabolomics knowledge graphs. The system enables collaborative exploration of spectra, annotations, taxa, and biological activities within a KG-driven workflow.
Graph-of-Skills proposes dependency-aware structural retrieval for massive agent skill libraries to address token budget and hallucination in retrieval. It retrieves skills with awareness of prerequisite chains, enabling scalable and context-appropriate skill selection for agents.
This work tackles open-world interactive object search in household environments by modeling relational semantics over 3D scene graphs. They propose SCOUT, which builds scene graphs and learns a utility function to guide exploration, capturing relationships between objects and context. The approach aims to outperform vision-language embedding methods and large language models in real-time, context-aware search.
This work addresses reliance on isolated visual evidence in medical image classification by proposing a case-aware reasoning framework driven by multimodal knowledge graphs. It connects images with similar historical cases and their symptoms to build evidence-aware reasoning, and uses reliability-guided refinement to improve predictions. The goal is more accurate and interpretable clinical diagnosis.
This study investigates whether neuroscience expert-level reasoning can emerge from knowledge graphs built from a single authoritative textbook. It constructs a high-quality KG from the textbook and evaluates KG-grounded question-answering to probe deep reasoning capabilities without relying on external corpora. The results explore the potential of structured knowledge to enable advanced reasoning in neuroscience.
This paper proposes a mixture-of-experts knowledge graph retrieval-augmented generation framework for multi-agent LLM-based recommendations. It addresses varying query complexity and knowledge granularity by routing queries to specialized KG experts and grounding generation in up-to-date KG evidence. The approach aims to improve recommendation quality and reliability in multi-agent settings.
AlayaLaser is an on-disk, graph-based vector similarity search system that shows compute-bound bottlenecks at high dimensionality. It introduces an efficient index layout and search strategy to reduce compute overhead and improve throughput for large-scale high-dimensional vector retrieval.