Edge Graph Research Articles

A Fast Algorithm for All-Pairs-Shortest-Paths Suitable for Neural Networks.

Given a directed graph of nodes and edges connecting them, a common problem is to find the shortest path between any two nodes. Here we show that the shortest path distances can be found by a simple matrix inversion: if the edges are given by the adjacency matrix Aij, then with a suitably small value of γ, the shortest path distances are Dij=ceil(logγ[(I-γA)-1]ij).We derive several graph-theoretic bounds on the value of γ and explore its useful range with numerics on different graph types. Even when the distance function is not globally accurate across the entire graph, it still works locally to instruct pursuit of the shortest path. In this mode, it also extends to weighted graphs with positive edge weights. For a wide range of dense graphs, this distance function is computationally faster than the best available alternative. Finally, we show that this method leads naturally to a neural network solution of the all-pairs-shortest-path problem.

An efficient water quality index forecasting and categorization using optimized Deep Capsule Crystal Edge Graph neural network.

The world's freshwater supply, predominantly sourced from rivers, faces significant contamination from various economic activities, confirming that the quality of river water is critical for public health, environmental sustainability, and effective pollution control. This research addresses the urgent need for accurate and reliable water quality monitoring by introducing a novel method for estimating the water quality index (WQI). The proposed approach combines cutting-edge optimization techniques with Deep Capsule Crystal Edge Graph neural networks, marking a significant advancement in the field. The innovation lies in the integration of a Hybrid Crested Porcupine Genghis Khan Shark Optimization Algorithm for precise feature selection, ensuring that the most relevant indicators of water quality (WQ) are utilized. Furthermore, the use of the Greylag Goose Optimization Algorithm to fine-tune the neural network's weight parameters enhances the model's predictive accuracy. This dual optimization framework significantly improves WQI prediction, achieving a remarkable mean squared error (MSE) of 6.7 and an accuracy of 99%. By providing a robust and highly accurate method for WQ assessment, this research offers a powerful tool for environmental authorities to proactively manage river WQ, prevent pollution, and evaluate the success of restoration efforts. PRACTITIONER POINTS: Novel method combines optimization and Deep Capsule Crystal Edge Graph for WQI estimation. Preprocessing includes data cleanup and feature selection using advanced algorithms. Deep Capsule Crystal Edge Graph neural network predicts WQI with high accuracy. Greylag Goose Optimization fine-tunes network parameters for precise forecasts. Proposed method achieves low MSE of 6.7 and high accuracy of 99%.

Emergency voltage control strategy for power system transient stability enhancement based on edge graph convolutional network reinforcement learning

Emergency control is essential for maintaining the stability of power systems, serving as a key defense mechanism against the destabilization and cascading failures triggered by faults. Under-voltage load shedding is a popular and effective approach for emergency control. However, with the increasing complexity and scale of power systems and the rise in uncertainty factors, traditional approaches struggle with computation speed, accuracy, and scalability issues. Deep reinforcement learning holds significant potential for the power system decision-making problems. However, existing deep reinforcement learning algorithms have limitations in effectively leveraging diverse operational features, which affects the reliability and efficiency of emergency control strategies. This paper presents an innovative approach for real-time emergency voltage control strategies for transient stability enhancement through the integration of edge-graph convolutional networks with reinforcement learning. This method transforms the traditional emergency control optimization problem into a sequential decision-making process. By utilizing the edge-graph convolutional neural network, it efficiently extracts critical information on the correlation between the power system operation status and node branch information, as well as the uncertainty factors involved. Moreover, the clipped double Q-learning, delayed policy update, and target policy smoothing are introduced to effectively solve the issues of overestimation and abnormal sensitivity to hyperparameters in the deep deterministic policy gradient algorithm. The effectiveness of the proposed method in emergency control decision-making is verified by the IEEE 39-bus system and the IEEE 118-bus system.

Analytical approach to piezoelectric model synthesis with the use of Cauer’s method for system design

Background Piezoceramic materials have unique property which enables direct and bilateral conversion between mechanical and electrical energy. This ability facilitates significant miniaturisation of technology and opens many opportunities in design of new actuators and energy harvesters. Mathematical modelling of piezoelectric modules is notoriously hard due to complex constitutive equations defining mechanical and electrical energy conversion. Methods The article presents research on a new synthesis method based on the Cauer’s method. Mechanical damping is introduced with the use of Rayleigh’s approximation. A discrete electromechanical model is formed based on the Mason’s piezoelectric model. The proposed approach allows modelling of piezoelectric systems based on a set of characteristic frequencies. The method allows a more general approach to the problem of modelling new systems, as opposed to application-oriented methods seen in literature. A non-standard model analysis method using edge graphs and structural numbers is also verified as a potential alternative for matrix-based method. The authors compare their precision and computation requirements. Results The structural method of mechanical model analysis gave identical results as the reference matrix method. However, the non-classical algorithm took much longer to calculate and was using more memory. The electromechanical model analysis has shown an error of 5% in comparison to resonance frequencies taken from a reference plate specification. The calculated magnitude of displacement was well above the capability of a 3.5mm thick piezoelectric plate. Conclusions The synthesis method presented in this paper allows synthesizing piezoelectric cascade models based on limited information in form of characteristic frequencies. Currently this method allows a coarse approximation of the real piezoelectric parameters with limited number of inputs. The additional method of analysis based on structural numbers offers a promising alternative to matrix calculations but requires a more thorough investigation of the computational power required to determine whether it can compete with existing algorithms.

Graph semantic information for self-supervised monocular depth estimation

Self-supervised monocular depth estimation has garnered significant attention in recent years due to its practical value in applications, as it eliminates the need for ground truth depth maps during training. However, its performance usually drops when estimating weakly textured regions and boundary regions, primarily due to the limited depth representation capability of traditional Convolutional Neural Networks (CNNs) that do not support topology. To address these issues, we propose a Graph Semantic Model (GSM) to improve self-supervised monocular depth estimation by utilizing graph learning and semantic information. Our focus is on improving feature representation through graph semantic information. Therefore, we incorporate semantic segmentation and depth estimation into one framework and enhance the interaction of different modal information through the Inter-Directed Graph Reasoning (IDGR) module. In addition, we design the Semantic-Guided Edge Graph Reasoning (SGEGR) module, aiming to boost the network's ability to perceive local depth. Extensive experiments on the KITTI dataset show that our method outperforms the state-of-the-art methods, particularly in accurately estimating depth within weakly textured regions and boundary regions.

Hypergraph Structural Information Aggregation Generative Adversarial Networks for Diagnosis and Pathogenetic Factors Identification of Alzheimer's Disease With Imaging Genetic Data.

Alzheimer's disease (AD) is a neurodegenerative disease with profound pathogenetic causes. Imaging genetic data analysis can provide comprehensive insights into its causes. To fully utilize the multi-level information in the data, this article proposes a hypergraph structural information aggregation model, and constructs a novel deep learning method named hypergraph structural information aggregation generative adversarial networks (HSIA-GANs) for the automatic sample classification and accurate feature extraction. Specifically, HSIA-GAN is composed of generator and discriminator. The generator has three main functions. First, vertex graph and edge graph are constructed based on the input hypergraph to present the low-order relations. Second, the low-order structural information of hypergraph is extracted by the designed vertex convolution layers and edge convolution layers. Finally, the synthetic hypergraph is generated as the input of the discriminator. The discriminator can extract the high-order structural information directly from hypergraph through vertex-edge convolution, fuse the high and low-order structural information, and finalize the results through the full connection (FC) layers. Based on the data acquired from AD neuroimaging initiative, HSIA-GAN shows significant advantages in three classification tasks, and extracts discriminant features conducive to better disease classification.

Applying self-supervised learning to network intrusion detection for network flows with graph neural network

Graph Neural Networks (GNNs) have garnered intensive attention for Network Intrusion Detection System (NIDS) due to their suitability for representing the network traffic flows. However, most present GNN-based methods for NIDS are supervised or semi-supervised. Network flows need to be manually annotated as supervisory labels, a process that is time-consuming or even impossible, making NIDS difficult to adapt to potentially complex attacks, especially in large-scale real-world scenarios. The existing GNN-based self-supervised methods focus on the binary classification of network flow as benign or not, and thus fail to reveal the types of attack in practice. This paper studies the application of GNNs to identify the specific types of network flows in an unsupervised manner. We first design an encoder to obtain graph embedding, that introduces the graph attention mechanism and considers the edge information as the only essential factor. Then, a self-supervised method based on graph contrastive learning is proposed. The method samples center nodes, and for each center node, generates subgraph by it and its direct neighbor nodes, and corresponding contrastive subgraph from the interpolated graph, and finally constructs positive and negative samples from subgraphs. Furthermore, a structured contrastive loss function based on edge features and graph local topology is introduced. To the best of our knowledge, it is the first GNN-based self-supervised method for the multiclass classification of network flows in NIDS. Detailed experiments conducted on four real-world datasets (NF-Bot-IoT, NF-Bot-IoT-v2, NF-CSE-CIC-IDS2018, and NF-CSE-CIC-IDS2018-v2) systematically compare our model with the state-of-the-art supervised and self-supervised models, illustrating the considerable potential of our method. Our code is accessible through https://github.com/renj-xu/NEGSC.

Analytical approach to piezoelectric model synthesis with the use of Cauer's method for system design.

Piezoceramic materials have unique property which enables direct and bilateral conversion between mechanical and electrical energy. This ability facilitates significant miniaturisation of technology and opens many opportunities in design of new actuators and energy harvesters. Mathematical modelling of piezoelectric modules is notoriously hard due to complex constitutive equations defining mechanical and electrical energy conversion. The article presents research on a new synthesis method based on the Cauer's method. Mechanical damping is introduced with the use of Rayleigh's approximation. A discrete electromechanical model is formed based on the Mason's piezoelectric model. The proposed approach allows modelling of piezoelectric systems based on a set of characteristic frequencies. The method allows a more general approach to the problem of modelling new systems, as opposed to application-oriented methods seen in literature. A non-standard model analysis method using edge graphs and structural numbers is also verified as a potential alternative for matrix-based method. The authors compare their precision and computation requirements. The structural method of mechanical model analysis gave identical results as the reference matrix method. However, the non-classical algorithm took much longer to calculate and was using more memory. The electromechanical model analysis has shown an error of 5% in comparison to resonance frequencies taken from a reference plate specification. The calculated magnitude of displacement was well above the capability of a 3.5mm thick piezoelectric plate. The synthesis method presented in this paper allows synthesizing piezoelectric cascade models based on limited information in form of characteristic frequencies. Currently this method allows a coarse approximation of the real piezoelectric parameters with limited number of inputs. The additional method of analysis based on structural numbers offers a promising alternative to matrix calculations but requires a more thorough investigation of the computational power required to determine whether it can compete with existing algorithms.

Reinforce crystal material property prediction with comprehensive message passing via deep graph networks

Machine learning methods have gained extensive attention in the field of material design and discovery. Graph neural networks (GNN) have shown promise in predicting of material properties, particularly within the context of crystal materials, which naturally lend themselves to graph representations. In these representations, atoms and their corresponding bonds serve as nodes and edges within the graph. However, most existing GNN models focus on capturing the structure–property relationship using atomic species and bond length, neglecting crucial structural information inherent to crystal materials. Notably, the bond angle is a critical structural parameter. In this study, we propose a novel model, the Crystal Gated Graph Attention Network (CGGAT), designed for the precise prediction of crystal material properties. The CGGAT model combines the Graph Attention Network (GAT) and Gated Graph ConvNets (GatedGCN), which transfer messages between atom graphs and edge graphs that characterize the crystal structure. The GAT layer aggregates information from the local neighborhood of each node to extract important information in the atom and edge graphs. We trained and tested the model on the Materials Project and JARVIS-DFT databases. The experimental results demonstrate that CGGAT can predict crystal material properties more quickly than density functional theory and surpasses other machine learning model algorithms in the same prediction tasks.

Multi-ship encounter situation graph structure learning for ship collision avoidance based on AIS big data with spatio-temporal edge and node attention graph convolutional networks

With the increasing number of ships on the sea, the frequency multi-ship encounters situation was becoming more common than two-ship encounter. The complexity and risk of the navigation will exponentially increase with the more ships involved. Relying solely on the International Regulations for Preventing Collisions at Sea (COLREGS) (objective knowledge) was insufficient to handle the multi-ship intelligent collision avoidance problem, also needed the ship officer's good seamanship (subjective knowledge). In this study, we propose a methodology that combines subjective insights from AIS big data with objective analysis through multi-ship encounters recognition with graph convolutional networks (GCN). (1) The ship encounter 8-azimuths map was utilized to identify the two-ship encounter situation (25 types) from the AIS data. (2) Identify the multi-ship encounters trajectory data by cross-matching the two-ship encounter data. (3) To handle the intricate relationship information between the multiple ships which were transformed into a graph structure using graph theory. (4) Finally, a spatial-temporal edge and node attention graph convolutional network (ST-ENAGCN) was proposed with the graph convolutional unit and long short-term memory (LSTM) unit. The 2022 Ningbo Sea area AIS big data was utilized to achieve graph-structured learning regarding human experiences during the multi-ship encounters situation. The results indicate that the ST-ENAGCN model can understand the complex marine traffic situation to make the own ship collision avoidance decisions. This study contributes significantly to the increased efficiency and safety of sea operations in complex marine traffic conditions and support autonomous navigation of swarm of MASSs under the human-machine hybrid (HMH) conditions to better understand the collision avoidance intention of the ship officer in the multi-ship encounters situation.

Stability theory for two-lobe states on the tadpole graph for the NLS equation

The aim of this work is to present new spectral tools for studying the orbital stability of standing waves solutions for the nonlinear Schrödinger equation (NLS) with power nonlinearity on a tadpole graph, namely, a graph consisting of a circle with a half-line attached at a single vertex. By considering δ-type boundary conditions at the junction and bound states with a positive two-lobe profile, the main novelty of this paper is at least twofold. Via a splitting eigenvalue method developed by the author, we identify the Morse index and the nullity index of a specific linearized operator around of an a priori positive two-lobe state profile for every positive power; and we also obtain new results about the existence and the orbital stability of positive two-lobe states at least in the cubic NLS case. To our knowledge, the results contained in this paper are the first in studying positive bound states for the NLS on a tadpole graph by non-variational techniques. In particular, our approach has prospect of being extended to study stability properties of other bound states for the NLS on a tadpole graph or on other non-compact metric graph such as a looping edge graph, as well as, for other nonlinear evolution models on a tadpole graph.

Assessing electrogenetic activation via a network model of biological signal propagation

Introduction: Molecular communication is the transfer of information encoded by molecular structure and activity. We examine molecular communication within bacterial consortia as cells with diverse biosynthetic capabilities can be assembled for enhanced function. Their coordination, both in terms of engineered genetic circuits within individual cells as well as their population-scale functions, is needed to ensure robust performance. We have suggested that “electrogenetics,” the use of electronics to activate specific genetic circuits, is a means by which electronic devices can mediate molecular communication, ultimately enabling programmable control.Methods: Here, we have developed a graphical network model for dynamically assessing electronic and molecular signal propagation schemes wherein nodes represent individual cells, and their edges represent communication channels by which signaling molecules are transferred. We utilize graph properties such as edge dynamics and graph topology to interrogate the signaling dynamics of specific engineered bacterial consortia.Results: We were able to recapitulate previous experimental systems with our model. In addition, we found that networks with more distinct subpopulations (high network modularity) propagated signals more slowly than randomized networks, while strategic arrangement of subpopulations with respect to the inducer source (an electrode) can increase signal output and outperform otherwise homogeneous networks.Discussion: We developed this model to better understand our previous experimental results, but also to enable future designs wherein subpopulation composition, genetic circuits, and spatial configurations can be varied to tune performance. We suggest that this work may provide insight into the signaling which occurs in synthetically assembled systems as well as native microbial communities.

Context Matters: Distilling Knowledge Graph for Enhanced Object Detection

The human visual system is capable of not only recognizing individual objects but also comprehending the contextual relationship between them in real-world scenarios, making it highly advantageous for object detection. However, in practical applications, such contextual information is often not available. Previous attempts to compensate for this by utilizing cross-modal data such as language and statistics to obtain contextual priors have been deemed sub-optimal due to a semantic gap. To overcome this challenge, we present a seamless integration of context into an object detector through Knowledge Distillation. Our approach intuitively represents context as a knowledge graph, describing the relative location and semantic relevance of different visual concepts. Leveraging recent advancements in graph representation learning with Transformer, we exploit the contextual information among objects using edge encoding and graph attention. Specifically, each image region propagates and aggregates the representation from its highly similar neighbors to form the knowledge graph in the Transformer encoder. Extensive experiments and a thorough ablation study conducted on challenging benchmarks MS-COCO, Pascal VOC and LVIS demonstrate the superiority of our method.

Modern graphic design application innovation based on graphic abstract processing method

Abstract The art of abstraction reflects the rich imagination of human beings to a certain extent and is now widely used in movies, games, advertisements, and other modern design fields. In this paper, a new edge tangential flow algorithm applicable to image abstraction is proposed, which first estimates the gradient strength and gradient direction of each point in a given input image by using the Sobel operator and then establishes the edge graph of the input image by using the edge information transfer strategy. On this basis, the application and value of abstract art and this graphical abstract processing method in modern graphic design are discussed, and the essence of abstract form art is perfectly embodied so as to make the two perfectly integrated. The results show that in the application of graphic abstract processing in graphic design, the average value of user attraction, understanding, and memory dimension scores are 0.53, 0.675, and 0.683, respectively, which verifies the effectiveness of this graphic design practice through users.

Exploratory Search Prompt Generation using n-Degree Connection in Knowledge Graph

Search engines play a vital role in retrieving in- formation, but users often struggle to express their precise information needs, resulting in less-than-optimal search results. Therefore, enhancing search query refinement is crucial to elevate the accuracy and relevance of search outcomes. One particular challenge that existing search engines face is presenting refined results for queries containing two or more unrelated entities.
 This paper presents a novel approach for efficient search prompt generation by leveraging connected nodes and attributes in the knowledge graph. We propose a comprehensive exploration technique that explores the n-degrees connections and their at- tributes to generate all possible imaginative prompts. We realized that n-degrees connection exploration is an expensive task, hence we conducted experiments to determine the effectiveness of 2- degree exploration prompts in covering all the user-asked queries within the provided dataset.
 In testing with approximately 2000 queries on a related knowl- edge graph dataset, we found out that our proposed methodol- ogy significantly improved query coverage. At n=1 depth, the coverage increased from 58% without the methodology to 84% with it. At n=2 depth, the coverage rose from 92% without the methodology to nearly 99% with it. Additionally, due to our question caching strategy at n=2, we observed faster response times for all questions.

Balancing graph Voronoi diagrams with one more vertex

AbstractLet be a graph with unit‐length edges and nonnegative costs assigned to its vertices. Given a list of pairwise different vertices , the prioritized Voronoi diagram of with respect to is the partition of in subsets so that, for every with , a vertex is in if and only if is a closest vertex to in and there is no closest vertex to in within the subset . For every with , the load of vertex equals the sum of the costs of all vertices in . The load of equals the maximum load of a vertex in . We study the problem of adding one more vertex at the end of in order to minimize the load. This problem occurs in the context of optimally locating a new service facility (e.g., a school or a hospital) while taking into account already existing facilities, and with the goal of minimizing the maximum congestion at a site. There is a brute‐force algorithm for solving this problem in time on ‐vertex ‐edge graphs. We prove a matching time lower bound–up to sub‐polynomial factors–for the special case where and , assuming the so called Hitting Set Conjecture of Abboud et al. On the positive side, we present simple linear‐time algorithms for this problem on cliques, paths and cycles, and almost linear‐time algorithms for trees, proper interval graphs and (assuming to be a constant) bounded‐treewidth graphs.

EGS P26 Facilitating the Emergency Surgical Ward Round: The Route to WardNav

Abstract Background The ward round is a corner stone of surgical practice, this is no different when dealing with the acute surgical take. Constant bed pressures result in frequently needing to visit multiple wards, requiring time and effort on staff to ensure patients are reviewed. With frequent rotation of junior doctors, often the responsibility of planning the ward round falls to the ward round leader, who is also making the clinical decisions. We hypothesise a ward round planning tool available to the surgical team would result in more efficient ward rounds, freeing the team to focus on clinical decisions. Methods A survey of junior doctors (foundation doctors and core trainees) was distributed assessing their level of engagement with ward round planning while on surgical rotations at Southmead Hospital, Bristol. An initial planning tool was created in Microsoft Excel utilising a node edge graph and an algorithm based on the Hamiltonian cycle. After further survey feedback, the tool was developed into a web-based app using a version of Djikstra’s algorithm. Results 25 junior doctors responded to the initial survey over two rotations. 100% had to go to multiple wards on a round, 64% actively think about the route required although only 28% would be confident in planning the route. 96% would find a planning tool helpful. Feedback on the Excel based tool from the first cohort revealed appetite for a mobile app-based version. The resultant application was created in R. The final version allows the user to start the round on any ward and creates the optimal route through the building, visually represented on a graph with accompanying text output. Conclusions Junior surgical doctors are not confident with planning ward rounds in a large teaching hospital. An app-based ward navigation tool has the potential to remove ward round planning from daily decision making and engage the junior surgical staff.

Analytical approach to piezoelectric model synthesis with the use of Cauer’s method for system design

Background: Piezoceramic materials have unique property which enables direct and bilateral conversion between mechanical and electrical energy. This ability facilitates significant miniaturisation of technology and opens many opportunities in design of new actuators and energy harvesters. Mathematical modelling of piezoelectric modules is notoriously hard due to complex constitutive equations defining mechanical and electrical energy conversion. Methods: The article presents research on a new synthesis method based on the Cauer’s method for electrical and mechanical system design. Mechanical damping is introduced with the use of Rayleigh’s approximation. A discrete electromechanical model is formed based on the Mason’s piezoelectric model. As an additional element of the research, a non-standard model analysis method using edge graphs and structural numbers is employed to investigate potential reductions in computational requirements for model analysis. Results: The electromechanical model analysis has shown an error of 5% in relation to the resonance target. The calculated peak in magnitude of displacement at the resonant frequency was well above the capability of a 3.5mm thick piezoelectric plate. The proposed non-standard method of analysis gave identical results in terms of obtained resonance frequencies. The calculated magnitude of vibrations was varying by 50 – 58%. Conclusions: The synthesis method presented in this paper allows an approximation of piezoelectric parameters of a real system based on the created mathematical model. Currently this method is subject to a considerable error in the determined vibration amplitudes of the system, but it allows a coarse approximation of the parameters while having a very limited number of input data. The additional method of analysis based on structural numbers offers a promising alternative to matrix calculations but requires a more thorough investigation of the computational power required to determine whether it can compete with existing algorithms.

Factorization and pseudofactorization of weighted graphs.

For unweighted graphs, finding isometric embeddings of a graph is closely related to decompositions of into Cartesian products of smaller graphs. When is isomorphic to a Cartesian graph product, we call the factors of this product a factorization of . When is isomorphic to an isometric subgraph of a Cartesian graph product, we call those factors a pseudofactorization of . Prior work has shown that an unweighted graph's pseudofactorization can be used to generate a canonical isometric embedding into a product of the smallest possible pseudofactors. However, for arbitrary weighted graphs, which represent a richer variety of metric spaces, methods for finding isometric embeddings or determining their existence remain elusive, and indeed pseudofactorization and factorization have not previously been extended to this context. In this work, we address the problem of finding the factorization and pseudofactorization of a weighted graph , where satisfies the property that every edge constitutes a shortest path between its endpoints. We term such graphs minimal graphs, noting that every graph can be made minimal by removing edges not affecting its path metric. We generalize pseudofactorization and factorization to minimal graphs and develop new proof techniques that extend the previously proposed algorithms due to Graham and Winkler [Graham and Winkler, '85] and Feder [Feder, '92] for pseudofactorization and factorization of unweighted graphs. We show that any -vertex, -edge graph with positive integer edge weights can be factored in time, plus the time to find all pairs shortest paths (APSP) distances in a weighted graph, resulting in an overall running time of time. We also show that a pseudofactorization for such a graph can be computed in time, plus the time to solve APSP, resulting in an running time.

Long-term multivariate time series forecasting in data centers based on multi-factor separation evolutionary spatial–temporal graph neural networks

Data center infrastructures require constant monitoring to ensure stable and reliable operation and time-series forecasting plays an indispensable role in intelligent operations and maintenance in data centers. However, the potential for accurate time-series predictions is often limited due to the overlooked relationships between data records from independent sensors. Inferring relationships for a potential graph representation of a data center is challenging due to complex relationships between nodes and multiple factors that may cause connections between them. Moreover, graphs change dynamically in long-term predictions, but current methods do not account for future graph changes. To address these challenges, we propose a long-term time-series forecasting framework called Multi-factor Separation Evolutionary Spatial–Temporal Graph Neural Networks (MSE-STGNN). Our framework considers edge diversity, graph changes and spatial–temporal architecture in long-term prediction processes and proposes three modules. Specifically, we propose a Multi-factor Separation (MS) module to separate the factors influencing node connectivity, enabling the acquisition of a graph more closely aligned with actual circumstances; then we propose a Graph Prediction (GP) module to incorporate future graphs to correct errors in the graph on which multi-step predictions depend. Moreover, we propose an Attention-enhanced Spatial–temporal dilated causal convolution module (AS-Conv) to more effectively leverage information pertaining to spatial and historical events. Our experimental results on datasets comprising of temperature and IT power data collected from real-world data centers show that the proposed method outperforms other advanced prediction methods in terms of prediction accuracy, and the learned latent graphs are explainable.