Graph Convolution Research Articles

Interpretable multi-graph convolution network integrating spatio-temporal attention and dynamic combination for wind power forecasting

Graph neural networks (GNN) can effectively improve wind power forecasting accuracy due to their ability to capture spatial correlations between wind farms. However, conventional GNN-based forecasting models constructing the graphs from only one specific perspective fail to characterize the spatial correlation patterns comprehensively and accurately. Furthermore, their black-box structure leads to deficient interpretability of the forecasting results, which hampers the quantification of forecasting uncertainty and reduces the credibility of applying these forecasting models in practice. To address the above problems, an interpretable ultra-short-term wind power forecasting method is proposed based on multi-graph convolution network integrating spatio-temporal attention and dynamic graph combination. Four undirected weight graphs are constructed to form a multi-graph scheme by considering various geographical and statistical aspects, and then graph convolution network (GCN) and temporal convolution network (TCN) are used to enhance the forecasting performance by leveraging information in both space and time dimensions. Meanwhile, it consists of temporal attention that deeply mines the temporal dependencies and spatial attention that interprets the influence of wind farms on the forecasts. Additionally, to boost the adaptability and generalization capability of the proposed method, an improved reinforcement learning is elaborately designed to combine the results of different graphs by dynamically optimizing their weights to obtain the final forecasts. Experiment results suggest that the proposed method exhibits superior forecasting accuracy and reasonable interpretability compared with other benchmark methods.

Multi‐scale skeleton simplification graph convolutional network for skeleton‐based action recognition

AbstractHuman action recognition based on graph convolutional networks (GCNs) is one of the hotspots in computer vision. However, previous methods generally rely on handcrafted graph, which limits the effectiveness of the model in characterising the connections between indirectly connected joints. The limitation leads to weakened connections when joints are separated by long distances. To address the above issue, the authors propose a skeleton simplification method which aims to reduce the number of joints and the distance between joints by merging adjacent joints into simplified joints. Group convolutional block is devised to extract the internal features of the simplified joints. Additionally, the authors enhance the method by introducing multi‐scale modelling, which maps inputs into sequences across various levels of simplification. Combining with spatial temporal graph convolution, a multi‐scale skeleton simplification GCN for skeleton‐based action recognition (M3S‐GCN) is proposed for fusing multi‐scale skeleton sequences and modelling the connections between joints. Finally, M3S‐GCN is evaluated on five benchmarks of NTU RGB+D 60 (C‐Sub, C‐View), NTU RGB+D 120 (X‐Sub, X‐Set) and NW‐UCLA datasets. Experimental results show that the authors’ M3S‐GCN achieves state‐of‐the‐art performance with the accuracies of 93.0%, 97.0% and 91.2% on C‐Sub, C‐View and X‐Set benchmarks, which validates the effectiveness of the method.

A fine-grained grading network for natural products based on dynamic association inference

Fine-grained grading of natural products, based on internal defects, plays an important role in food processing. Although X-ray imaging is critical for revealing internal defects due to its capacity to penetrate surfaces, the inherent grayscale and lower contrast make it challenging to extract fine-grained features. To address this, we propose the Fine-Grained Grading Network (FGGN) for classifying the quality grades of multiple natural products from X-ray images. Leveraging a ResNet18 backbone, FGGN enhances feature extraction by employing a hybrid attention module that emphasizes crucial local details by analyzing joint dependencies within tensor dimensions. Moreover, by combining cross-dimensional context encoding and dynamic graph convolution, a dynamic correlation inference module is proposed to construct dynamic graphs in a gated inference pattern, thereby integrating fine-grained information across abstraction levels. Finally, the cross-entropy loss is combined with graph constraint loss to train the network end-to-end. Extensive evaluation of three natural product datasets shows that the Fine-Grained Grading Network achieves state-of-the-art recognition performance, which demonstrates its grading applicability to multiple natural products.

Graph recurrent neural networks-integrated real-time prediction of key displacements for fire-induced collapse early warning of steel frames

This study proposes the FRAME-Net, which can extract spatial-temporal features of discrete structures like planar steel frames, for early warning systems of fire-induced building collapse. FRAME-Net, based on graph neural networks and recurrent neural networks, seamlessly integrates feature extraction techniques, advanced graph convolution operations, and an enhanced graph long short-term memory mechanism, resolving the critical issue that the actual state of burning buildings, e.g., fire scenario, load case, material properties, etc., cannot be identified accurately and rapidly. The temperatures of steel components and easy-to-measure displacements are inputs, and hard-to-measure displacements at the top and interior of steel frames under fire can be predicted in real time. Numerical examples indicate that the trained agent accurately predicts the hard-to-measure displacements for the trained steel frame under completely unknown datasets, showcasing its potential in real-world unknown fire scenarios. Besides, the trained intelligent agent can be directly applied to untrained steel frames with different topological forms without re-training, which significantly reduces computational costs and is beneficial for swift emergency responses. Subsequently, satisfactory results predicted by the trained agent without re-training for the steel frame under the real fire test further underscores the promising generalization capability of the trained agent to adapt to unknown fire scenarios in real-world steel frames. Finally, the feasibility of the proposed framework to facilitate the implementation of the early-warning systems of fire-induced collapse is proved by comparing the predicted and actual remaining building collapse time, thereby assisting in reducing casualties during the rescue process and contributing to smart firefighting advancements.

IC points weight learning-based GCN and improving feature distribution for industrial fault diagnosis

Industrial fault diagnosis (FD) is crucial to detect the causes of faults in a timely manner and solve the problems to maintain stable production. Adequately considering the spatial structure of the samples and their features is an urgent problem for better FD. This paper focused on the points located in the within-class boundaries and between-class intersection region (IC points) and proposed IC point weight learning-based graph convolution network (GCN) and improving feature distribution (ICWGCN-FD) method. Firstly, ICWGCN-FD assigns appropriate weights to IC points specially allocated by graph learning layer. Secondly, a cosine distance loss term is particularly designed to make the IC points closer to the center of their respective classes and further away from their nearest neighbor points belongs to other classes. Finally, the spatial feature distribution of the overall samples is improved under the influence of the message passing function possessed by GCN, thus presenting a clearer distribution. Two cases from industrial simulation are carried out to validate the practical feasibility and superiority of ICWGCN-FD, compared with other classic and recent graph neural networks. In addition, the 3D visualization results correspond to the principles of the method, and an ablation experiment successfully verified the validity of each component.

Graph Adaptive Attention Network with Cross-Entropy.

Non-Euclidean data, such as social networks and citation relationships between documents, have node and structural information. The Graph Convolutional Network (GCN) can automatically learn node features and association information between nodes. The core ideology of the Graph Convolutional Network is to aggregate node information by using edge information, thereby generating a new node feature. In updating node features, there are two core influencing factors. One is the number of neighboring nodes of the central node; the other is the contribution of the neighboring nodes to the central node. Due to the previous GCN methods not simultaneously considering the numbers and different contributions of neighboring nodes to the central node, we design the adaptive attention mechanism (AAM). To further enhance the representational capability of the model, we utilize Multi-Head Graph Convolution (MHGC). Finally, we adopt the cross-entropy (CE) loss function to describe the difference between the predicted results of node categories and the ground truth (GT). Combined with backpropagation, this ultimately achieves accurate node classification. Based on the AAM, MHGC, and CE, we contrive the novel Graph Adaptive Attention Network (GAAN). The experiments show that classification accuracy achieves outstanding performances on Cora, Citeseer, and Pubmed datasets.

KG-LIME: predicting individualized risk of adverse drug events for multiple sclerosis disease-modifying therapy.

The aim of this project was to create time-aware, individual-level risk score models for adverse drug events related to multiple sclerosis disease-modifying therapy and to provide interpretable explanations for model prediction behavior. We used temporal sequences of observational medical outcomes partnership common data model (OMOP CDM) concepts derived from an electronic health record as model features. Each concept was assigned an embedding representation that was learned from a graph convolution network trained on a knowledge graph (KG) of OMOP concept relationships. Concept embeddings were fed into long short-term memory networks for 1-year adverse event prediction following drug exposure. Finally, we implemented a novel extension of the local interpretable model agnostic explanation (LIME) method, knowledge graph LIME (KG-LIME) to leverage the KG and explain individual predictions of each model. For a set of 4859 patients, we found that our model was effective at predicting 32 out of 56 adverse event types (P < .05) when compared to demographics and past diagnosis as variables. We also assessed discrimination in the form of area under the curve (AUC = 0.77 ± 0.15) and area under the precision-recall curve (AUC-PR = 0.31 ± 0.27) and assessed calibration in the form of Brier score (BS = 0.04 ± 0.04). Additionally, KG-LIME generated interpretable literature-validated lists of relevant medical concepts used for prediction. Many of our risk models demonstrated high calibration and discrimination for adverse event prediction. Furthermore, our novel KG-LIME method was able to utilize the knowledge graph to highlight concepts that were important to prediction. Future work will be required to further explore the temporal window of adverse event occurrence beyond the generic 1-year window used here, particularly for short-term inpatient adverse events and long-term severe adverse events.

Dynamic community detection algorithm based on hyperbolic graph convolution

PurposeCommunity detection is a key factor in analyzing the structural features of complex networks. However, traditional dynamic community detection methods often fail to effectively solve the problems of deep network information loss and computational complexity in hyperbolic space. To address this challenge, a hyperbolic space-based dynamic graph neural network community detection model (HSDCDM) is proposed.Design/methodology/approachHSDCDM first projects the node features into the hyperbolic space and then utilizes the hyperbolic graph convolution module on the Poincaré and Lorentz models to realize feature fusion and information transfer. In addition, the parallel optimized temporal memory module ensures fast and accurate capture of time domain information over extended periods. Finally, the community clustering module divides the community structure by combining the node characteristics of the space domain and the time domain. To evaluate the performance of HSDCDM, experiments are conducted on both artificial and real datasets.FindingsExperimental results on complex networks demonstrate that HSDCDM significantly enhances the quality of community detection in hierarchical networks. It shows an average improvement of 7.29% in NMI and a 9.07% increase in ARI across datasets compared to traditional methods. For complex networks with non-Euclidean geometric structures, the HSDCDM model incorporating hyperbolic geometry can better handle the discontinuity of the metric space, provides a more compact embedding that preserves the data structure, and offers advantages over methods based on Euclidean geometry methods.Originality/valueThis model aggregates the potential information of nodes in space through manifold-preserving distribution mapping and hyperbolic graph topology modules. Moreover, it optimizes the Simple Recurrent Unit (SRU) on the hyperbolic space Lorentz model to effectively extract time series data in hyperbolic space, thereby enhancing computing efficiency by eliminating the reliance on tangent space.

Multi-attention gated temporal graph convolution neural Network for traffic flow forecasting

Multi-attention gated temporal graph convolution neural Network for traffic flow forecasting

Phase-wise attention GCN for recommendation denoising

In recent years, Graph Convolution Network (GCN) has been widely applied in recommendation systems. Compared to traditional collaborative filtering methods that only aggregate information of neighboring nodes, GCN-based recommendation systems can aggregate information of high-order neighborhoods of a node to achieve better performance. However, GCN-based recommendation systems have faced the over-smoothing problem. The over-smoothing problem causes the performance of recommender systems first to rise and then decrease as the number of stacked layers increases. Previous GCN-based recommender systems have provided a few solutions to over-smoothing. We believe that there are still two problems: (1) most models focus on aggregating diverse information, with less consideration of filtering noise; (2) most models ignore the fact that there are differences in different stages of the process of aggregating information using GCN. We introduce a Phase-wise Attention mechanism to address these issues and propose a Phase-wise Attention GCN (PAGCN) recommendation model. Considering the different characteristics of different stages during the aggregation process of the GCN-based recommendation systems, our model uses different information aggregation methods in different graph convolution layers and adopts targeted aggregation methods. In this way, high-order neighborhood information can be more controlled to improve the performance and effectiveness of our model. The experimental results on real-world datasets show that our model outperforms various baselines, demonstrating the reasonableness of our method.

A human activity recognition method based on Vision Transformer

Human activity recognition has a wide range of applications in various fields, such as video surveillance, virtual reality and human–computer intelligent interaction. It has emerged as a significant research area in computer vision. GCN (Graph Convolutional networks) have recently been widely used in these fields and have made great performance. However, there are still some challenges including over-smoothing problem caused by stack graph convolutions and deficient semantics correlation to capture the large movements between time sequences. Vision Transformer (ViT) is utilized in many 2D and 3D image fields and has surprised results. In our work, we propose a novel human activity recognition method based on ViT (HAR-ViT). We integrate enhanced AGCL (eAGCL) in 2s-AGCN to ViT to make it process spatio-temporal data (3D skeleton) and make full use of spatial features. The position encoder module orders the non-sequenced information while the transformer encoder efficiently compresses sequence data features to enhance calculation speed. Human activity recognition is accomplished through multi-layer perceptron (MLP) classifier. Experimental results demonstrate that the proposed method achieves SOTA performance on three extensively used datasets, NTU RGB+D 60, NTU RGB+D 120 and Kinetics-Skeleton 400.

Composite Neighbor-Aware Convolutional Metric Networks for Hyperspectral Image Classification.

Supervised classification of hyperspectral image (HSI) is generally required to obtain better performance in spectral-spatial feature learning by fully using complex pixel- and superpixel-level interdependencies with small labeled samples. Limited by the local regular convolutions, convolutional neural networks (CNNs) can only exploit information from the short-range Euclidean neighbors of a target, hindering the effectiveness of feature representation. In contrast, graph convolutional networks (GCNs) can learn long-range dependencies between non-Euclidean neighbors but usually require the input of a full graph constructed from a whole HSI, making GCNs must be trained in a full-batch manner with tremendous computational consumption. In this work, we propose a composite neighbor-aware convolutional metric network (CNCMN), aiming to learn each target's representation from its composite neighbors (i.e., both Euclidean and non-Euclidean neighbors) in a batchwise manner. Specifically, for each target in an HSI, its Euclidean neighbors are the pixels in the local square region centered on itself, and its non-Euclidean neighbors are several related nodes selected from the constructed full graph. Correspondingly, a composite convolution (CoConv) is proposed by coupling an image convolution and a graph convolution, which can perform flexible convolutions on those composite neighbors and extract adaptively fused features from them. Besides, to further boost classification, we also propose a mini-batch metric classifier to dynamically optimize interclass and intraclass distances of samples batch by batch, which is then combined with the CoConv to form the mini-batch CNCMN. Extensive experiments on three real-world HSIs demonstrate the advantages of the proposed method over mini-batch deep learning algorithms and have obtained the state-of-the-art performance in these fields. The code is available at: https://github.com/qichaoliu/HSI-CNCMN.

CSDN: Cross-Modal Shape-Transfer Dual-Refinement Network for Point Cloud Completion.

How will you repair a physical object with some missings? You may imagine its original shape from previously captured images, recover its overall (global) but coarse shape first, and then refine its local details. We are motivated to imitate the physical repair procedure to address point cloud completion. To this end, we propose a cross-modal shape-transfer dual-refinement network (termed CSDN), a coarse-to-fine paradigm with images of full-cycle participation, for quality point cloud completion. CSDN mainly consists of "shape fusion" and "dual-refinement" modules to tackle the cross-modal challenge. The first module transfers the intrinsic shape characteristics from single images to guide the geometry generation of the missing regions of point clouds, in which we propose IPAdaIN to embed the global features of both the image and the partial point cloud into completion. The second module refines the coarse output by adjusting the positions of the generated points, where the local refinement unit exploits the geometric relation between the novel and the input points by graph convolution, and the global constraint unit utilizes the input image to fine-tune the generated offset. Different from most existing approaches, CSDN not only explores the complementary information from images but also effectively exploits cross-modal data in the whole coarse-to-fine completion procedure. Experimental results indicate that CSDN performs favorably against twelve competitors on the cross-modal benchmark.

Short-Term Power Forecasting of Wind Farm Cluster Based on Global Information Adaptive Perceptual Graph Convolution Network

Short-Term Power Forecasting of Wind Farm Cluster Based on Global Information Adaptive Perceptual Graph Convolution Network

Hierarchical Bidirected Graph Convolutions for Large-Scale 3-D Point Cloud Place Recognition.

In this article, we present a novel hierarchical bidirected graph convolution network (HiBi-GCN) for large-scale 3-D point cloud place recognition. Unlike place recognition methods based on 2-D images, those based on 3-D point cloud data are typically robust to substantial changes in real-world environments. However, these methods have difficulty in defining convolution for point cloud data to extract informative features. To solve this problem, we propose a new hierarchical kernel defined as a hierarchical graph structure through unsupervised clustering from the data. In particular, we pool hierarchical graphs from the fine to coarse direction using pooling edges and fuse the pooled graphs from the coarse to fine direction using fusing edges. The proposed method can, thus, learn representative features hierarchically and probabilistically; moreover, it can extract discriminative and informative global descriptors for place recognition. Experimental results demonstrate that the proposed hierarchical graph structure is more suitable for point clouds to represent real-world 3-D scenes.

Power grid network security: A lightweight detection model for composite false data injection attacks using spatiotemporal features

The stability of power systems is paramount to industrial operations. The deleterious inherent characteristics of false data injection attacks (FDIA) have drawn substantial interest due to their severe threats to power grids. Contemporary detection systems face numerous challenges as attackers employ various tactics, such as injecting complex elements into measurement data and formulating quick attack strategies against critical nodes and transmission lines in the power grid network topology. Conventional models often fail to adapt to the intricacies of practical situations because they focus predominantly on detecting individual components. To overcome the above predicaments, this paper proposes a lightweight detection model integrating deep separable convolutional layers, squeeze neural networks, and a bidirectional long short-term memory architecture named DSE-BiLSTM. The acquisition process of network topological characteristics is accomplished through variable graph attention autoencoder (VGAAE). This approach leverages the effectiveness of the graph convolution (GCN) layer to acquire each node’s topological feature and the graph attention (GAT) module to identify and extract the topological features of critical nodes. Furthermore, the topology information obtained by the both techniques is embedded in one-dimensional vector space in the same form as measurement data. By combining the output of VGAAE with meter measurements, the feature fusion of temporal and spatial modalities is realized. DSE-BiLSTM with optimal hyperparameters achieves an F1-score of 99.56% and a row accuracy (RACC) of 93.10% on the conventional dataset. The experimental results of FDIA detection with composite datasets of IEEE 14-bus and IEEE 118-bus systems show that the F1-score and RACC of DSE-BiLSTM remain above 84.51% and 83.56% under various attack strengths and noise levels. In addition, as the power grid network scales up, noise level’s effect on detection performance decreases, while attack strength’s effect on recognition capability increases. DSE-BiLSTM can effectively process the composite data of spatiotemporal multimodes and provides a feasible solution for the localization and detection of FDIA in realistic scenes.

Influential simplices mining via simplicial convolutional networks

The identification of influential simplices is crucial for understanding higher-order network dynamics. Yet, despite relatively mature research on influential nodes (0-simplices) mining, characterizing simplices’ influence and identifying influential simplices remain challenging due to notable discrepancies in vital nodes and vital simplices mining. In this paper, we propose a higher-order graph learning model, named influential simplices mining neural networks (ISMnet), to identify vital simplices in simplicial complexes. ISMnet leverages novel higher-order representations: hierarchical bipartite graphs and higher-order hierarchical (HoH) Laplacians, where target simplices are grouped into a hub set and can interact with other simplices. It also employs learnable graph convolution operators in each HoH Laplacian domain to capture interactions among simplices and can identify influential simplices of arbitrary order by changing the hub set. Notably, ISMnet addresses the limitations inherent in traditional graph neural networks that struggle with higher-order tasks, while seamlessly retaining the capability to exploit network topology and node features concurrently. Numerical results on empirical and synthetic datasets demonstrate that ISMnet significantly outperforms existing methods by at least 12% and 4%, respectively, in ranking 2-simplices. In general, this novel framework promises to serve as a potent tool in higher-order network analysis.

Multi-Interest Sequential Recommendation with Simplified Graph Convolution and Multiple Item Features

Multi-interest sequential recommendations leverage users’ historical behavior to provide recommendations that match multiple interests. Most of these methods have not fully extracted higher-order information hidden in users’ interactions and have overlooked the multiple features of items. To this end, this paper proposes a multi-interest model called “multi-interest sequential recommendation with simplified graph convolution and item multi-features (SGCMF)”. Firstly, a simplified graph convolution module is designed based on bipartite graphs, which utilizes mean pooling to aggregate neighboring information and employs a feedforward neural network (FNN) for nonlinear transformations and combinations. This method reduces redundant information and captures higher-order relationships, thereby simplifying the complexity of modeling high-order interactions and improving prediction accuracy. Secondly, an item multi-feature extraction module is proposed, which represents item features with multiple vectors, and analyzes each feature from multiple perspectives while preserving important relationships between features. The model correlates multiple features of the item with user interests, thereby achieving a fine-grained analysis of user interests. Extensive experiments are conducted on five real-world scenarios, and the results are compared with state-of-the-art methods. The experimental results show that SGCMF outperforms other baselines.

HOGFormer: high-order graph convolution transformer for 3D human pose estimation

HOGFormer: high-order graph convolution transformer for 3D human pose estimation

Decomposition with feature attention and graph convolution network for traffic forecasting

Traffic forecasting is a crucial task for enhancing the quality and efficiency of intelligent transportation systems. In recent years, several neural networks have been proposed to tackle this challenge. However, traffic data itself has complex composition and intertwined temporal patterns. Currently, the absence of effective methods to capture the inherent properties and overall profile of traffic data has become a bottleneck in improving the traffic prediction capabilities of models. To address these issues, we introduce a novel model, Decomposition with Feature Attention and Graph Convolution Network (DFAGCN), for Traffic Forecasting. To highlight the inherent properties of complicated traffic data, we present a unique Decomposition with Feature Attention (DFA) method to decompose the traffic data into seasonality, trend, and remainder. On the one hand, we introduce a new method to filter out noise and possible outliers from time series data. On the other hand, we employ an attention mechanism to make our decomposition method adaptively extract the trend. Subsequently, we utilize spatial–temporal attention to simulate the dynamic spatial–temporal correlations of traffic data. Finally, we use spatial–temporal convolution to extract local dependencies of traffic data. Experiments conducted on four real-world traffic forecasting datasets demonstrate that DFAGCN achieves state-of-the-art performance. The results show that the application of time series decomposition to traffic data significantly enhances the accuracy of traffic prediction.