Showing 14 papers for 2026-07-03
We propose X-LogSMask, an explainable multi-head logarithmic structural mask that injects symmetrically normalized graph topology directly into Transformer attention to better align with sparse, structured, and multi-scale graph data. This masking approach enhances interpretability while enabling efficient modeling of graph interactions.
We present MKGR, a multimodal representation framework for cold-start PPI prediction that combines region-aware protein sequence encoding with four protein-centered biomedical knowledge graphs. By fusing sequence information with graph contexts, MKGR improves predictive performance for proteins with no observed interactions during training.
We propose EHHN, an event-driven heterogeneous hypergraph network for object-centric next-activity prediction. It uses OCELs and a heterogeneous hypergraph to capture multi-object, cross-type interactions beyond pairwise relations, enabling better prediction.
We propose SABER, a semantic-aligned brain network analysis framework that actively integrates LLM-derived semantics into brain connectivity analysis using multi-scale hypergraphs. The method injects ROI-level semantics into prediction to improve diagnosis robustness and accuracy.
We introduce SA-HGNN, a Sample-Adaptive Hyperbolic Graph Neural Network for EEG-based depression recognition. By modeling brain connectivity in hyperbolic space and adapting to sample-specific hierarchies, SA-HGNN better captures hierarchical structure and improves classification performance.
We propose an additive MLP-GNN framework for predicting aqueous solubility that keeps physicochemical descriptors and molecular graph topology in separate branches: a descriptor encoder (MLP) and a graph encoder (GNN). The outputs are combined additively to enable attribution of solubility to global chemistry or structure.
We present a high-dimensional change point detection method via a graph-spanning ratio, applicable to offline and online settings and to Euclidean or graph-structured data with unknown distributions. The method provides error control and achieves near-optimal power when changes exceed the minimax separation rate.
We provide a practical evaluation of deep learning models and LLMs for vulnerability detection, examining their performance under real-world data distributions and label noise. Our study shows that gains on curated benchmarks often degrade in the wild, highlighting gaps and offering guidelines for robust deployment.
We introduce geometry-aware graph fusion for rainfall field reconstruction, addressing heterogeneous supports (points, paths, grids) by preserving each source's geometric constraints rather than merging in feature space. The method yields accurate, high-resolution rainfall fields by integrating multi-source data with geometry-consistent fusion.
We map text to a multiplex graph and formulate prompt compression as redundancy-aware graph pruning guided by Levy-walk dynamics. The approach preserves informative dependencies while reducing prompt length.
We propose robust and explainable 3D mode shape recognition with region-aware GNNs and a canonical engineering graph representation, improving accuracy across different vehicle architectures and FE meshes. The method provides interpretable reasoning for NVH analyses.
We present Scene-Conditioned PINN-GNN for multipath RF maps, enabling cross-scene generation and in-scene completion with 2D and 2.5D representations. By combining physics-informed neural networks with graph neural networks, the framework enforces electromagnetic propagation laws while leveraging spatial structure.
We introduce CoPersona, collaborative persona graphs for robust LLM personalization. By building a graph of user personas and borrowing information from related users, CoPersona fills sparse user profiles to enable more reliable personalization.
When RAG Meets Query Planning proposes using logical query trees to plan exploration in exploratory reasoning problems, addressing retrieval noise and error accumulation. The approach borrows ideas from database query planning to create end-to-end plans that guide retrieval and reasoning paths.