Showing 15 papers for 2026-06-04
We introduce In-Context Graphical Inference (ICG-I), an autoregressive Graph Transformer designed to preserve the sequential elimination structure used in exact marginal inference. By aligning the inductive bias of a neural model with the exact inference workflow, ICG-I aims to deliver scalable yet correct marginal inference for discrete graphical models, mitigating convergence issues on frustrated topologies.
Graph Cascades proposes a mesoscopic rewiring strategy for GNNs and Graph Transformers based on contagion diffusion that constructs an auxiliary graph in O(|V|+|E|). Node pairs reinforced by repeated multi-hop signals are promoted to direct neighbors, enabling the model to capture intermediate-scale structure beyond local edges or global attention. The authors characterize when such reinforcement-based rewiring is beneficial and provide sufficient conditions.
We introduce Graph Set Transformer (GST), a neural network for learning on sets of graphs where per-element predictions depend on both local structure and set-wide context. Unlike prior methods that require pre-encoded graph embeddings, GST interleaves node-level feature propagation with cross-graph contextualization at each layer, fusing the two levels of information.
SpliceBind is a graph neural network framework for isoform-aware druggability prediction, addressing splice-mediated drug resistance across isoforms. It improves prediction accuracy (AUROC 0.703 vs. 0.634 for baseline P2Rank) and analyzes when structure-based methods succeed or fail across six clinically validated variants spanning different isoforms.
DegradoMap is a graph neural network that predicts PROTAC-mediated degradability using only protein structure and E3 ligase identity, i.e., at the target-selection stage. It encodes biophysical priors into the model to reflect degradability determinants and works without a complete PROTAC structure.
We introduce Bayesian Membership Privacy (BMP), a sampling-aware privacy notion for graph neural networks that accounts for graph structural correlations and stochastic training-graph sampling. Unlike prior node-prior privacy analyses, BMP incorporates node-dependent priors and treats graph-sampling probabilities as part of the privacy mechanism, enabling better privacy-utility tradeoffs.
EpiFormer uses geometric deep learning to learn antigen-antibody interactions for epitope prediction. It addresses two fundamental challenges: encoding antibody chains jointly rather than independently and dealing with class imbalance and scarce antibody-antigen complexes. The model captures co-dependent structural features to improve binding interface predictions.
HYolo integrates hypergraph learning into YOLO to capture higher-order object-context relationships beyond pairwise interactions. This augmentation enables richer contextual modeling and improves object representation. Experiments on COCO show significant performance gains.
Identifying and Correcting Label Noise for Robust GNNs via Influence Contradiction introduces ICGNN that leverages graph structure to detect and rectify noisy labels through influence contradictions, improving robustness of GNNs under annotation noise.
Fixed Aggregation Features (FAFs) challenge the view that trainable neighborhood aggregations are essential for GNNs. FAFs convert graph tasks into tabular problems using fixed aggregation features and enable strong, training-free performance with well-tuned tabular classifiers, rivaling or surpassing many GNNs.
This study investigates what structural inductive biases help transformers reason over knowledge graphs, via controlled ablations of sparse adjacency masking, edge-type biases, query scaling, and value gating. The key finding is that sparse adjacency masking alone accounts for most of the gains in multi-hop reasoning across benchmarks, while the other components contribute modestly.
SFMambaNet proposes a Spectral-Frequency Enhanced Selective State Space Model for correspondence pruning, addressing the limitation that Mamba-based methods accumulate inconsistent features in the state space. The model leverages spectral-frequency features to better distinguish inliers from outliers in correspondences.
Treat Traffic Like Trees presents a semantic-preserving hierarchical graph-based expert framework for encrypted traffic analysis. It emphasizes preserving protocol semantics and leverages the hierarchical structure of protocol layers to analyze encrypted traffic more interpretably and effectively.
Generalizing Graph Foundation Models via Hyperbolic Retrieval-Augmented Generation proposes a retrieval-augmented generation framework in hyperbolic space to better handle hierarchical graph data and distribution shifts. The approach aims to generalize GFMs by leveraging hyperbolic geometry for knowledge retrieval, improving cross-domain robustness.
Breaking the Likelihood Trap: Consistent Generative Recommendation with Graph-structured Model proposes a generative reranking framework that mitigates the likelihood trap—repetitive high-likelihood sequences—by enforcing consistency constraints or exploiting graph-structured dependencies, yielding more coherent and engaging recommendations.