Showing 15 papers for 2026-03-09
We investigate how graph-based encoding improves NBA salary prediction over standard tabular features by modeling relationships among players, teams, and performance context. The graph-based approach yields better predictions, particularly for veterans or outliers, and provides insights into when relational encoding helps.
We introduce STEP, a framework that reframes temporal link prediction as a sequential forecasting task in continuous time, capturing evolving topology and temporal motif transitions. It treats future events as a sequence of stochastic motifs, improving prediction over standard binary link prediction with negative sampling. Experiments on real-world timestamped networks show improved accuracy and uncertainty calibration.
This work proposes a stock price forecasting framework that fuses a node transformer architecture with BERT-based sentiment analysis to capture cross-sectional dependencies and market sentiment. The integrated model leverages structural relationships among stocks and textual signals to improve forecasting accuracy over baselines.
We study ensemble learning for probabilistic sea surface temperature forecasting using Graph Neural Networks, focusing on how input perturbations affect forecast skill and uncertainty representation. A homogeneous bagging-like ensemble is implemented for the Canary Islands region, illustrating how perturbation design shapes predictive performance and uncertainty estimates.
This paper introduces polarized direct cross-attention (PolaDCA) for graph neural networks to enable adaptive, data-driven message passing in machinery fault diagnosis. By relaxing static graphs and homogeneous aggregation, PolaDCA improves robustness to changing relational structures in fault data.
We propose Expert-Aided Causal Discovery of Ancestral Graphs, enabling ex-ante incorporation of background knowledge and ex-post refinement via expert queries. The approach aims to improve reliability when causal sufficiency is violated by leveraging targeted expert input to refine causal graphs.
We provide a geometric explanation for the difficulties of learning GNN-based SAT solvers using graph Ricci curvature. The results show that bipartite graphs derived from random k-SAT formulas exhibit inherently negative curvature, and this curvature diminishes solver performance on harder instances.
LLMTM benchmarks and optimizes LLMs for temporal motif analysis in dynamic graphs. The work systematically evaluates LLM capabilities on motif-related tasks and provides guidance for effective deployment of LLMs in temporal graph analysis.
The EpisTwin presents a knowledge graph-grounded neuro-symbolic architecture for personal AI. It grounds generative reasoning in a verifiable Personal Knowledge Graph and leverages multimodal language models to fuse heterogeneous, cross-application data.
Towards Neural Graph Data Management introduces NGDBench, a unified benchmark to evaluate neural graph databases across domains, supporting the full Cypher query language for complex pattern matching beyond basic operations.
SCOUT tackles open-world interactive object search by using relational reasoning on 3D scene graphs. It guides exploration with a learned utility function over relations, addressing limitations of vision-language embeddings and large language models.
MASFactory offers a graph-centric framework to orchestrate LLM-based multi-agent systems. By modeling workflows as directed graphs, it enables reusable, adaptable agent collaboration and easier integration of heterogeneous context sources.
A causal-graph approach to oppositional narrative analysis models narratives as entity-level causal graphs to detect, analyze, and classify opposing narratives, offering a structured alternative to ontology-dependent text analysis.
Knowledge graphs are presented as implicit reward models: path-derived signals enable compositional reasoning. The paper proposes a post-training pipeline combining supervised fine-tuning and reinforcement learning, where knowledge graphs guide reasoning.
AutothinkRAG introduces complexity-aware control of retrieval-augmented reasoning for image-text interaction. It addresses long-context challenges and reasoning bottlenecks in vision-language models by adaptively managing retrieval and reasoning complexity.