Towards the Explainability of Temporal Graph Networks via Memory Backtracking and Topological Attribution

CATGX: Causal-Aware Temporal Graph Explanation via Scalable Motif Sampling and Adjustment

Experiments demonstrate that CATGX strikes a good balance to generate explanations that are faithful, grounded and confounder-aware, and outperform existing TGNN explainers in efficiency with 83x speed up.

CausalST: directed delay-aware graph neural networks for traffic flow prediction

Much of the recent work on traffic forecasting relies on two implicit simplifications: representing the road network as an undirected graph and adopting temporally symmetric encoders within the historical observation window. These choices may obscure asymmetric predictive relations among sensors and the chronological ordering of intermediate temporal representations. CausalST addresses these limitations jointly. For spatial modeling, separate source and destination embeddings are assigned to each sensor to construct a learned directed adjacency under road-topology constraints. On the temporal side, dilated causal convolutions serve as a temporal inductive bias, ensuring that the representation at each historical position depends only on current and earlier observations. Delay-aware graph propagation aggregates current and lagged neighbor states over predefined discrete lags. A parallel spatial attention branch captures nonlocal dependencies beyond the physical topology, while a gated fusion unit adaptively integrates the temporal, graph-propagation, and spatial-attention representations. Experiments on PeMS03, PeMS04, and PeMS08 show that CausalST achieves the lowest MAPE among all compared methods while remaining competitive in MAE and RMSE. Ablation results reveal non-additive interactions between directed weighting and delayed aggregation, with the full configuration attaining the lowest MAE on both evaluated datasets.

Temporal Knowledge Graph Forecasting via Constrained Large Language Models

Predicting Probabilistic Links Using Node Embedding and Graph Neural Network Approach

Link prediction describes the procedure for discovering out how likely it is that an edge will be present in a graph network. In the realm of graph network and graph data analytics, link prediction is regarded as a crucial challenge. Estimating possible future linkages among nodes in a network using an overview of the network at a certain moment in time is the focus of link prediction. This technique is utilized in various domains, including recommended systems, network traffic prediction, protein-protein interactions, and numerous other areas. Typically, conventional methods are implemented to quantify the degree of similarity between two specified nodes in a fixed network. The evaluation of networks that undergo dynamic changes over time, however, has recently become the subject of academic research. Heuristic node similarity scores, latent node embeddings, and explicit node attributes are the mainstays of current link prediction methods. Over time, graph neural networks (GNNs) have proven themselves superior to more conventional techniques of link prediction due to their ability to concurrently learn from graph structure and node/edge properties. A graph neural network model for predicting probabilistic links with network embedding via generalized graph convolutional networks, referred to as LPNE2GGCN for graph network datasets, has been proposed. This approach draws inspiration from the promising outcomes achieved by GCNs in evaluating such data. The LPNE2GGCN approach leverages node embedding (Node2Vec) via generalized graph convolutional networks (GraphSAGE). To assess its effectiveness, LPNE2GGCN has been compared against established latent node representation methods, including Node. The performance of LPNE2GGCN is evaluated by contrasting it with well-known techniques for latent node representation. The framework further improves link prediction accuracy by utilizing learning-based optimization techniques like Adadelta and Adam. The experimental results show that the suggested strategy outperforms the state-of-the-art across multiple social network datasets, including those using node embedding techniques and various approaches based on graph neural networks. The proposed model is based on graph neural network techniques that integrates the features obtained from node2vec and utilizes them as input for a hidden neural network consisting of three layers. The initial results suggest that this methodology produces excellent results compared to conventional methods across diverse social network datasets.

Structure-Guided Spatiotemporal Attention Graph Neural Network for Traffic Flow Prediction

The Structure-Guided Spatiotemporal Attention Graph Neural Network is proposed, offering a mechanistic account of the model's decision-making process while ensuring robust forecasting by aligning attention-based reasoning with identified macroscopic dependencies and preventing over-reliance on ephemeral local noise.

Faithful, Sufficient and Understandable: Rethinking Graph Counterfactual Explanations via Discrete Diffusion Inversion

This work proposes Graph Diffusion Counterfactual Explanation via Inversion (GDCE-I), a discrete denoising diffusion model with a novel discrete inversion scheme that enables distribution-aware edits leveraging the whole domain edit space and qualitatively shows that GDCE-I attains interpretable in-distribution solutions.