TodyNet: Temporal dynamic graph neural network for multivariate time series classification

Traffic forecasting plays an important role in intelligent transportation systems. However, the prediction task is highly challenging due to the mixture of global and local spatiotemporal dependencies involved in traffic data. Existing graph neural networks (GNNs) typically capture spatial dependencies with the predefined or learnable static graph structure, ignoring the hidden dynamic patterns in traffic networks. Meanwhile, most recurrent neural networks (RNNs) or convolutional neural networks (CNNs) cannot effectively capture temporal correlations, especially for long-term temporal dependencies. In this paper, we propose a spatial–temporal attention graph convolution network (STAGCN), which acquires a static graph and a dynamic graph from data without any prior knowledge. The static graph aims to model global space adaptability, and the dynamic graph is designed to capture local dynamics in the traffic network. A gated temporal attention module is further introduced for long-term temporal dependencies, where a causal-trend attention mechanism is proposed to increase the awareness of causality and local trends in time series. Extensive experiments on four real-world traffic flow datasets demonstrate that STAGCN achieves an outstanding prediction accuracy improvement over existing solutions.

Graph Neural Networks (GNNs) have become the de facto standard for representation learning on topological graphs, which usually derive effective node representations via message passing from neighborhoods. Although GNNs have achieved great success, previous models are mostly confined to static and homogeneous graphs. However, there are multiple dynamic interactions between different-typed nodes in real-world scenarios like academic networks and e-commerce platforms, forming temporal heterogeneous graphs (THGs). Limited work has been done for representation learning on THGs and the challenges are in two aspects. First, there are abundant dynamic semantics between nodes while traditional techniques like meta-paths can only capture static relevance. Second, different semantics on THGs have often mutually evolved with each other over time, making it more difficult than dynamic homogeneous graph modeling. To address this problem, here we propose the Dynamic Meta-path guided temporal heterogeneous Graph Neural Networks (DyMGNNs). To handle the dynamic semantics, we introduce the concept of dynamic meta-path which is a common base for temporal semantic search engines, and then adopt the temporal importance sampling to extract neighborhoods with temporal bias. Focusing on mutual evolution, we design the heterogeneous mutual evolution attention mechanism, which can model the fine-grained interplay of semanticlevel preferences for each node. Extensive experiments on three real-world datasets for node classification and temporal link prediction demonstrate that our method consistently outperforms state-of-the-art alternatives.

Spatial temporal incidence dynamic graph neural networks for traffic flow forecasting

A spatial–temporal graph neural network framework for automated software bug triaging

Dynamic recommendation systems aim to provide personalized suggestions by modeling temporal user-item interactions across time-series behavioral data. Recent studies have leveraged pre-trained dynamic graph neural networks (GNNs) to learn user-item representations over temporal snapshot graphs. However, fine-tuning GNNs on these graphs often results in generalization issues due to temporal discrepancies between pre-training and fine-tuning stages, limiting the model’s ability to capture evolving user preferences. To address this, we propose TarDGR, a task-aware retrieval-augmented framework designed to enhance generalization capability by incorporating task-aware model and retrieval-augmentation. Specifically, TarDGR introduces a Task-Aware Evaluation Mechanism to identify semantically relevant historical subgraphs, enabling the construction of task-specific datasets without manual labeling. It also presents a Graph Transformer-based Task-Aware Model that integrates semantic and structural encodings to assess subgraph relevance. During inference, TarDGR retrieves and fuses task-aware subgraphs with the query subgraph, enriching its representation and mitigating temporal generalization issues. Experiments on multiple large-scale dynamic graph datasets demonstrate that TarDGR consistently outperforms state-of-the-art methods, with extensive empirical evidence underscoring its superior accuracy and generalization capabilities.

Measures of centrality of vertices in a network are usually defined solely on the basis of the network structure. In highly dynamic networks, where vertices appear and disappear and their connectivity constantly changes, we need to redefine our measures of centrality to properly capture the temporal dimension of the network structure evolution, as well as the dynamic processes over the network. Betweenness centrality (BC), one of the most studied measures, defines the importance of a vertex as a mediator between available communication paths. BC value of a node is expressed as the fraction of the shortest paths passing through this node. In other words, given the context that information flow follows the shortest paths, a node with higher BC potentially has greater influence on the information flow in the network. In temporal dynamic graphs, a communication path should be seen as a path both in space (i.e., the network structure) and in time (i.e., the network evolution). Toward this goal, in this paper we propose the bi-objective notion of shortest–fastest path (SFP) in temporal graphs, which considers both space and time as a linear combination governed by a parameter. Based on this notion, we then define a novel temporal betweenness centrality (TBC) metric, which is highly sensitive to the observation interval and the importance of space and time distances of the vertices, that can provide better understanding of the communication mediators in temporal networks. We devise efficient algorithms for exactly computing all-pairs SFPs and the corresponding BC values both in a single graph window and sliding graph windows. We also present a distributed implementation of our approach on Apache Spark which shows great solution effectiveness and efficiency. We provide a thorough experimentation on a large variety of datasets. An application to the analysis of information propagation proves that our notion of TBC outperforms static BC in the task of identifying the best vertices for propagating information.

Link prediction in Online Social Networks—OSNs—has been the focus of numerous studies in the machine learning community. A successful machine learning-based solution for this task needs to (i) leverage global and local properties of the graph structure surrounding links; (ii) leverage the content produced by OSN users; and (iii) allow their representations to change over time, as thousands of new links between users and new content like textual posts, comments, images and videos are created/uploaded every month. Current works have successfully leveraged the structural information but only a few have also taken into account the textual content and/or the dynamicity of network structure and node attributes. In this paper, we propose a methodology based on temporal graph neural networks to handle the challenges described above. To understand the impact of textual content on this task, we provide a novel pipeline to include textual information alongside the structural one with the usage of BERT language models, dense preprocessing layers, and an effective post-processing decoder. We conducted the evaluation on a novel dataset gathered from an emerging blockchain-based online social network, using a live-update setting that takes into account the evolving nature of data and models. The dataset serves as a useful testing ground for link prediction evaluation because it provides high-resolution temporal information on link creation and textual content, characteristics hard to find in current benchmark datasets. Our results show that temporal graph learning is a promising solution for dynamic link prediction with text. Indeed, combining textual features and dynamic Graph Neural Networks—GNNs—leads to the best performances over time. On average, the textual content can enhance the performance of a dynamic GNN by 3.1% and, as the collection of documents increases in size over time, help even models that do not consider the structural information of the network.

DyHDGE: Dynamic heterogeneous transaction graph embedding for safety-centric fraud detection in financial scenarios

Traffic prediction has attracted a lot of attention in recent years. However, it is challenging due to the dynamic and heterogeneous spatial–temporal correlations. Most of the existing methods adopt the attention mechanism to capture the dynamic spatial features, but it is difficult for the attention mechanism to learn the real-time change of spatial dependencies, which restrict accurate spatial dependencies learning. Furthermore, most methods are out at elbows when solving heterogeneous spatial–temporal data. To overcome these problems, we propose a novel traffic prediction model called Dynamic spatial–temporal Heterogeneous Graph Convolution Network. Different from the existing methods, we have designed a dynamic localized graph and a corresponding adaptive localized graph convolution network, which are capable of simultaneously capturing dynamic and heterogeneous spatial correlations. We propose a gated adaptive temporal convolution network to capture the temporal heterogeneity of traffic data and enjoy global receptive fields. Finally, a global correlations fusion network is provided to incorporate the global spatial–temporal correlations. Compared with 11 baselines, our proposed model achieves state-of-the-art performance in the accuracy of prediction.

Evolving interdependencies across institutions and markets drive systemic financial risk, yet most forecasting models either treat assets independently or rely on static correlation structures. This limitation becomes particularly salient as cryptocurrency markets increasingly interact with traditional banking systems amid financial stress. Ignoring time-varying cross-market network structure risks understating tail risk ‎precisely during periods when accurate systemic risk assessment is most critical. This study proposes a dynamic graph neural network ‎‎(GNN) framework for systemic risk forecasting that models time-varying financial networks spanning banking institutions and major cryptocurrency assets. Nodes represent financial entities, while edges are constructed using rolling-window dependency measures that adapt to ‎changing market conditions. Node dynamics are modeled through temporal neural architectures, and stress regimes are explicitly identified ‎to evaluate performance under market turmoil. The empirical design includes strong temporal baselines, static-graph ablations, and cross-market removal experiments to isolate the contribution of network dynamics and crypto-market integration. Results indicate that a strong ‎LSTM baseline achieves superior volatility forecasting accuracy in both overall and stress-period evaluations, demonstrating the competitiveness of purely temporal models. However, within the class of graph-based models, dynamic GNNs substantially outperform static-graph variants during stress periods, demonstrating the importance of time-varying network structure for capturing volatility amplification. ‎Bank-only and full-system dynamic GNNs exhibit comparable stress-period performance, suggesting that cryptocurrency assets contribute ‎limited incremental information to bank-specific forecasts, while remaining informative for system-level stress characterization. The findings ‎suggest that dynamic graph representations enhance stress sensitivity and structural interpretability relative to static network models, even ‎when they do not surpass strong temporal baselines in raw predictive accuracy. The results support a restrained view of crypto–banking ‎contagion, emphasizing its conditional relevance during periods of market stress rather than unconditional systemic dominance.

Anomaly Detection in Dynamic Networks plays a critical role in various real-world applications such as cyberse-curity, e-commerce, and social media. Recent approaches based on graph neural networks have achieved fruitful results in static networks, and most of the researches focus on learning node embeddings to represent the large-scale complex networks, thereby facilitating downstream anomaly detection task. However, these methods require the whole network to extract intrinsic properties, so they are sensitive to the frequent changes of nodes and incapable of capturing the dynamism as realworld networks evolve over time. In this paper, we propose a novel framework DynAD for anomalous edge detection on time-evolving networks, which performs adaptive parameter learning in an end-to-end manner. In particular, DynAD first extracts fixed-size node embedding from each snapshot with temporal graph convolution and pooling operations. Then, it captures the temporal information of the graph sequence using Gated recurrent units (GRU) for anomaly detection, where an attention mechanism is employed to highlight the differences between the dynamic patterns. Experimental results on three real-world datasets illustrate that DynAD significantly outperforms the state-of-the-art baseline methods in anomaly detection.

Accurately forecasting the popularity of online information is critical for optimizing content delivery, recommendation systems, and network resource allocation. This paper introduces a novel framework that leverages temporal persistence patterns and graph neural networks (GNNs) to improve the prediction of information popularity. By modeling user-content interactions as dynamic graphs and incorporating historical popularity trends, our approach captures both structural and temporal dependencies. Extensive experiments on real-world social and content-sharing platforms demonstrate that the proposed method significantly outperforms traditional forecasting models in terms of accuracy and robustness. The results highlight the potential of combining graph-based learning with temporal analysis for intelligent information propagation modeling.

With information explosion occurring in past decades, the rapid growth of papers published results in the rapid change of hot topics, especially in the biomedical domain. It turns out very hard for researchers who are interested in biomedical domain to track hot topics over time, as well as to predict the trends of them in the near future. Based on the above demand, it is important to have a model which is able to follow and predict the trend of hot topics continuously. Deep learning has been proven to be an efficient method to extract information from texts and use the information to predict the future trends. Under the thriving background of Deep Learning, Graph Neural Network (GNN) is able to capture the information from graph structures. There are various applications using GNN models, such as traffic flow prediction, chemical structure discovering, etc. In this research project, a dynamic spatio-temporal graph neural network is presented to keep track of the selected hot keywords and topics in the biomedical domain and predict the possible frequencies in the near future. The input of the model is obtained by extracting the monthly frequency information of selected keywords and topics from paper abstracts in PubMed, the largest biomedical literature collection. After training with data over a decade, the model is able to predict trends of selected hot keywords and topics in next 5 months. Thus, the presented model can help follow the trend of hot topics in the biomedical domain.

Time series classification algorithms have been mainly dominated by non-deep learning models. Deep learning for Multivariate Time Series Classification (MTSC) has gained huge interest in recent years. Most state of the art deep learning methods are convolutional-based where 1-dimensional (1D) convolutions are used to extract features from the 2-dimensional time series. This study shows that factorization of 1D convolution filters into disjoint temporal and spatial components yields significant accuracy improvements with almost no additional computational cost. Based on our study on disjoint temporal-spatial filters, we have designed a novel filter block called "1+1D", which emphasizes the interaction between dimensions to improve the model performance of the convolution based on deep learning MTSC models. We also proposed a new and effective MTSC method called Disjoint-CNN using our proposed 1+1D filter blocks and through our extensive experiments show that our model (called Disjoint-CNN) outperforms the state-of-the-art MTSC models on 26 datasets in the UEA Multivariate time series archive, achieving the highest average rank among 9 MTSC benchmark models.

This paper proposes a physically constrained dynamic attention graph neural network (PC-DAGNN). Its core innovations lie in: first, constructing a physical prior constraint module to transform experimental theoretical formulas into hard-constrained loss terms, ensuring that the output conforms to physical laws; second, designing a dynamic attention graph neural network module to capture the coupling relationship between error sources at different experimental stages in real time. To verify the algorithm's performance, 10,000 sets of samples (8,000 training sets and 2,000 test sets, categorized as low, medium, and high error strength) containing random and systematic errors were generated using MATLAB, using the photoelectric effect and free fall experiments as experimental scenarios. The results were compared with the least squares (LS) method, a BP neural network, and a generalized graph neural network (GNN). Results show that the PC-DAGNN achieves a maximum reduction of 66% in mean absolute error (MAE) compared to the comparison algorithms, a consistently high physical law compliance rate (PCR) above 96%, and a generalization error fluctuation (GEF) only 50%-60% of that of the comparison algorithms. Research has demonstrated that PC-DAGNN can achieve high-precision and highly stable error analysis while ensuring physical logic consistency, providing an effective technical solution for error correction in complex physics experiments.