Over-smoothing Problem Research Articles

DSE-HNGCN: predicting the frequencies of drug-side effects based on heterogeneous networks with mining interactions between drugs and side effects.

Evaluating the frequencies of drug-side effects is crucial in drug development and risk-benefit analysis. While existing deep learning methods show promise, they have yet to explore using heterogeneous networks to simultaneously model the various relationship between drugs and side effects, highlighting areas for potential enhancement. In this study, we propose DSE-HNGCN, a novel method that leverages heterogeneous networks to simultaneously model the various relationships between drugs and side effects. By employing multi-layer graph convolutional networks, we aim to mine the interactions between drugs and side effects to predict the frequencies of drug-side effects. To address the over-smoothing problem in graph convolutional networks and capture diverse semantic information from different layers, we introduce a layer importance combination strategy. Additionally, we have developed an integrated prediction module that effectively utilizes drug and side effect features from different networks. Our experimental results, using benchmark datasets in a range of scenarios, show that our model outperforms existing methods in predicting the frequencies of drug-side effects. Comparative experiments and visual analysis highlight the substantial benefits of incorporating heterogeneous networks and other pertinent modules, thus improving the accuracy of DSE-HNGCN predictions. We also provide interpretability for DSE-HNGCN, indicating that the extracted features are potentially biologically significant. Case studies validate our model's capability to identify potential side effects of drugs, offering valuable insights for subsequent biological validation experiments.

ES-GNN: Generalizing Graph Neural Networks Beyond Homophily With Edge Splitting.

While Graph Neural Networks (GNNs) have achieved enormous success in multiple graph analytical tasks, modern variants mostly rely on the strong inductive bias of homophily. However, real-world networks typically exhibit both homophilic and heterophilic linking patterns, wherein adjacent nodes may share dissimilar attributes and distinct labels. Therefore, GNNs smoothing node proximity holistically may aggregate both task-relevant and irrelevant (even harmful) information, limiting their ability to generalize to heterophilic graphs and potentially causing non-robustness. In this work, we propose a novel Edge Splitting GNN (ES-GNN) framework to adaptively distinguish between graph edges either relevant or irrelevant to learning tasks. This essentially transfers the original graph into two subgraphs with the same node set but complementary edge sets dynamically. Given that, information propagation separately on these subgraphs and edge splitting are alternatively conducted, thus disentangling the task-relevant and irrelevant features. Theoretically, we show that our ES-GNN can be regarded as a solution to a disentangled graph denoising problem, which further illustrates our motivations and interprets the improved generalization beyond homophily. Extensive experiments over 11 benchmark and 1 synthetic datasets not only demonstrate the effective performance of ES-GNN but also highlight its robustness to adversarial graphs and mitigation of the over-smoothing problem.

SICGNN: Structurally informed convolutional graph neural networks for protein classification

Abstract Recently, graph neural networks (GNNs) have been widely used in various domains, including social networks, recommender systems, protein classification, molecular property prediction, and genetic networks. In bioinformatics and chemical engineering, considerable research is being actively conducted to represent molecules or proteins on graphs by conceptualizing atoms or amino acids as nodes and the relationships between nodes as edges. The overall structures of proteins and their interconnections are crucial for predicting and classifying their properties. However, as GNNs stack more layers to create deeper networks, the embeddings between nodes may become excessively similar, causing an oversmoothing problem that reduces the performance for downstream tasks. To avoid this, GNNs typically use a limited number of layers, which leads to the problem of reflecting only the local structure and neighborhood information rather than the global structure of the graph. Therefore, we propose a structurally informed convolutional GNN (SICGNN) that utilizes information that can express the overall topological structure of a protein graph during GNN training and prediction. By explicitly including information of the entire graph topology, the proposed model can utilize both local neighborhood and global structural information. We applied the SICGNN to representative GNNs such as GraphSAGE, graph isomorphism network, and graph attention network, and confirmed performance improvements across various datasets. We also demonstrate the robustness of SICGNN using multiple stratified 10-fold cross-validations and various hyperparameter settings, and demonstrate that its accuracy is comparable or better than those of existing GNN models.

Integrating contrastive learning and adversarial learning on graph denoising encoder for recommendation

In this era of information explosion, recommender systems have become increasingly vital in research and everyday applications. Collaborative filtering and matrix factorization are foundational techniques in this domain, while they encounter limitations in capturing complex user behavior and managing sparse data. Graph Convolutional Networks (GCNs) have emerged as a superior alternative. However, challenges for GCNs remain, particularly regarding data sparsity and interaction noise. We propose a Contrastive Adversarial Graph Denoising Encoder (CAGDE), which conducts contrastive learning and adversarial training into a Graph Denoising Encoder (GDE) framework to address these issues. This approach eliminates the necessity for neighborhood aggregation to address the over-smoothing problem and focuses on essential graph features by adding different types of noise to enhance robustness. Extensive experiments were conducted on five benchmark datasets, and the proposed solution, CAGDE, demonstrates significant improvements in recommendation accuracy and robustness against noise. Our implementations are available at https://github.com/Azunyan-bot/CAGDE.

Discrete Cosine Transform-Based Joint Spectral–Spatial Information Compression and Band-Correlation Calculation for Hyperspectral Feature Extraction

Prediction tasks over pixels in hyperspectral images (HSI) require careful effort to engineer the features used for learning a classifier. However, the generated classification map may suffer from an over-smoothing problem, which is manifested in significant differences from the original image in terms of object boundaries and details. To address this over-smoothing problem, we designed a method for extracting spectral–spatial-band-correlation (SSBC) features. In SSBC features, joint spectral–spatial feature extraction is considered a discrete cosine transform-based information compression, where a flattening operation is used to avoid the high computational cost induced by the requirement of distillation from 3D images for joint spectral–spatial information. However, this process can yield extracted features with lost spectral information. We argue that increasing the spectral information in the extracted features is the key to addressing the over-smoothing problem in the classification map. Consequently, the normalized difference vegetation index and iron oxide are improved for HSI data in extracting band-correlation features as added spectral information because their calculations, involving two spectral bands, are not appropriate for the abundant spectral bands of HSI. Experimental results on four real HSI datasets show that the proposed features can significantly mitigate the over-smoothing problem, and the classification performance is comparable to that of state-of-the-art deep features.

B cell epitope prediction by capturing spatial clustering property of the epitopes using graph attention network

Knowledge of B cell epitopes is critical to vaccine design, diagnostics, and therapeutics. As experimental validation for epitopes is time-consuming and costly, many in silico tools have been developed to computationally predict the B cell epitopes. While most methods show poor performance, deep learning methods in recent years have shown promising results. We developed a method called EpiGraph that outperformed previous methods, including those that showed a significant improvement in performance in recent years. Our model’s performance can be attributed to the following factors: (1) a combination of structure and sequence feature embeddings obtained from pretrained ESM-IF1 and ESM-2 models could capture the structural and evolutionary features of B cell epitopes, (2) a graph attention network could learn the spatial proximity of B cell epitopes with high graph homophily, and (3) residual connections in the model framework mitigate the over-smoothing problem in the graph neural network. Our model achieved the highest performance on an independent benchmark dataset. The results were also consistent on a different dataset. The datasets and source codes are available at https://github.com/sj584/EpiGraph. A user-friendly web server is freely available at http://epigraph.kaist.ac.kr.

Heterogeneous propagation graph convolution network for a recommendation system based on a knowledge graph

Recently, there has been a surge of interest in recommendation systems that leverage knowledge graphs, primarily because of their effectiveness in addressing sparsity and cold-start challenges inherent in collaborative filtering approaches. In most previous studies, researchers have focused on the way knowledge associations are encoded in knowledge graphs, but have not sufficiently highlighted the signals of collaboration that are implicit in the interaction between users and items. As a result, the learned embeddings do not provide a complete representation of the semantic information. In this paper, we describe a new model called a heterogeneous propagation graph convolution network for a recommendation system combined with a knowledge graph (HP-GCN). It adopts a heterogeneous propagation to generate user embedding representations, thereby combining encoded collaborative signals and auxiliary knowledge in knowledge graphs. Furthermore, we incorporate an attention mechanism to differentiate the contributions made by diverse neighbors as opposed to those made by users. Since most graph convolutions tend to suffer from over-smoothing when the number of convolutional layers increases, leading to insufficient utilization of high-order information, this paper uses an improved graph convolution strategy to generate item embeddings. This strategy has two different aggregation mechanisms embedding into different subgraphs, which can more fully utilize high-order information and mitigate the over-smoothing problem. Thus, we are able to efficiently prevent negative information originating from higher-order neighbors into the process of embedding learning. In extensive experiments, we applied HP-GCN to four large-scale real datasets for music, books, movies, and restaurants. The experimental outcomes revealed that HP-GCN generally surpassed the baseline methods in both recommendation accuracy and diversity, showing superior recommendation performance overall.

AGNN: Alternating Graph-Regularized Neural Networks to Alleviate Over-Smoothing.

Graph convolutional network (GCN) with the powerful capacity to explore graph-structural data has gained noticeable success in recent years. Nonetheless, most of the existing GCN-based models suffer from the notorious over-smoothing issue, owing to which shallow networks are extensively adopted. This may be problematic for complex graph datasets because a deeper GCN should be beneficial to propagating information across remote neighbors. Recent works have devoted effort to addressing over-smoothing problems, including establishing residual connection structure or fusing predictions from multilayer models. Because of the indistinguishable embeddings from deep layers, it is reasonable to generate more reliable predictions before conducting the combination of outputs from various layers. In light of this, we propose an alternating graph-regularized neural network (AGNN) composed of graph convolutional layer (GCL) and graph embedding layer (GEL). GEL is derived from the graph-regularized optimization containing Laplacian embedding term, which can alleviate the over-smoothing problem by periodic projection from the low-order feature space onto the high-order space. With more distinguishable features of distinct layers, an improved Adaboost strategy is utilized to aggregate outputs from each layer, which explores integrated embeddings of multi-hop neighbors. The proposed model is evaluated via a large number of experiments including performance comparison with some multilayer or multi-order graph neural networks, which reveals the superior performance improvement of AGNN compared with the state-of-the-art models.

EQGraphNet: Advancing single-station earthquake magnitude estimation via deep graph networks with residual connections

Magnitude estimation is a critical task in seismology, and conventional methods usually require dense seismic station arrays to provide data with sufficient spatiotemporal distribution. In this context, we propose the Earthquake Graph Network (EQGraphNet) to enhance the performance of single-station magnitude estimation. The backbone of the proposed model consists of eleven convolutional neural network layers and ten RCGL modules, where a RCGL combines a Residual Connection and a Graph convolutional Layer capable of mitigating the over-smoothing problem and simultaneously extracting temporal features of seismic signals. Our work uses the STanford EArthquake Dataset for model training and performance testing. Compared with three existing deep learning models, EQGraphNet demonstrates improved accuracy for both local magnitude and duration magnitude scales. To evaluate the robustness, we add natural background noise to the model input and find that EQGraphNet achieves the best results, particularly for signals with lower signal-to-noise ratios. Additionally, by replacing various network components and comparing their estimation performances, we illustrate the contribution of each part of EQGraphNet, validating the rationality of our approach. We also demonstrate the generalization capability of our model across different earthquakes occurring environments, achieving mean errors of ±0.1 units. Furthermore, by demonstrating the effectiveness of deeper architectures, this work encourages further exploration of deeper GNN models for both multi-station and single-station magnitude estimation.

Attention-based stackable graph convolutional network for multi-view learning

In multi-view learning, graph-based methods like Graph Convolutional Network (GCN) are extensively researched due to effective graph processing capabilities. However, most GCN-based methods often require complex preliminary operations such as sparsification, which may bring additional computation costs and training difficulties. Additionally, as the number of stacking layers increases in most GCN, over-smoothing problem arises, resulting in ineffective utilization of GCN capabilities. In this paper, we propose an attention-based stackable graph convolutional network that captures consistency across views and combines attention mechanism to exploit the powerful aggregation capability of GCN to effectively mitigate over-smoothing. Specifically, we introduce node self-attention to establish dynamic connections between nodes and generate view-specific representations. To maintain cross-view consistency, a data-driven approach is devised to assign attention weights to views, forming a common representation. Finally, based on residual connectivity, we apply an attention mechanism to the original projection features to generate layer-specific complementarity, which compensates for the information loss during graph convolution. Comprehensive experimental results demonstrate that the proposed method outperforms other state-of-the-art methods in multi-view semi-supervised tasks.

A graph residual generation network for node classification based on multi-information aggregation

The key to improving the performance of graph convolutional networks (GCN) is to fully explore the correlation between neighboring and distant information. Aiming at the over-smoothing problem of GCN, in order to make full use of the relationship among features, graphs and labels, a graph residual generation network based on multi-information aggregation (MIA-GRGN) is proposed. Firstly, aiming at the defects of GCN, we design a deep initial residual graph convolution network (DIRGCN), which connects the initial input through residuals, so that each layer node retains part of the information of the initial features, ensuring the localization of the graph structure and effectively alleviating the problem of over-smoothing. Secondly, we propose a random graph generation method (RGGM) by utilizing graph edge sampling and negative edge sampling, and optimize the supervision loss function of DIRGCN in the form of generation framework. Finally, applying RGGM and DIRGCN as inference modules for modeling hypotheses and obtaining approximate posterior distributions of unknown labels, an optimized loss function is obtained, we construct a multi-information aggregation MIA-GRGN that combines graph structure, node characteristics and label joint distribution. Experiments on benchmark graph classification datasets show that MIA-GRGN achieves better classification results compared with the benchmark models and mainstream models, especially for datasets with less dense edge relationships between nodes.

GTC: GNN-Transformer co-contrastive learning for self-supervised heterogeneous graph representation

Graph Neural Networks (GNNs) have emerged as the most powerful weapon for various graph tasks due to the message-passing mechanism’s great local information aggregation ability. However, over-smoothing has always hindered GNNs from going deeper and capturing multi-hop neighbors. Meanwhile, most methods follow a semi-supervised learning manner, the label scarcity would limit their applicability in real-world systems. Unlike GNNs, Transformers can model global information and multi-hop interactions via multi-head self-attention and a proper Transformer structure can show more immunity to over-smoothing. So, can we propose a novel framework to combine GNN and Transformer, integrating both GNN’s local information aggregation and Transformer’s global information modeling ability to eliminate the over-smoothing problem and achieve self-supervised graph representation? To realize this, this paper proposes a collaborative learning scheme for GNN-Transformer and constructs GTC architecture. GTC leverages the GNN and Transformer branch to encode node information from different views respectively, and establishes contrastive learning tasks based on the encoded cross-view information to realize self-supervised heterogeneous graph representation. For the Transformer branch, we propose Metapath-aware Hop2Token and CG-Hetphormer, which can cooperate with GNNs to attentively encode neighborhood information from different levels. As far as we know, this is the first attempt in the field of graph representation learning to utilize both GNNs and Transformer to collaboratively capture different view information and conduct cross-view contrastive learning. The experiments on real datasets show that GTC exhibits superior performance compared with state-of-the-art methods. Codes can be available at https://github.com/PHD-lanyu/GTC.

Multi-scale feature correspondence and pseudo label retraining strategy for weakly supervised semantic segmentation

Recently, the performance of semantic segmentation using weakly supervised learning has significantly improved. Weakly supervised semantic segmentation (WSSS) that uses only image-level labels has received widespread attention, it employs Class Activation Maps (CAM) to generate pseudo labels. Compared to traditional use of pixel-level labels, this technique greatly reduces annotation costs by utilizing simpler and more readily available image-level annotations. Besides, due to the local perceptual ability of Convolutional Neural Networks (CNN), the generated CAM cannot activate the entire object area. Researchers have found that this CNN limitation can be compensated for by using Vision Transformer (ViT). However, ViT also introduces an over-smoothing problem. Recent research has made good progress in solving this issue, but when discussing CAM and its related segmentation predictions, it is easy to overlook their intrinsic information and the interrelationships between them. In this paper, we propose a Multi-Scale Feature Correspondence (MSFC) method. Our MSFC can obtain the feature correspondence of CAM and segmentation predictions at different scales, re-extract useful semantic information from them, enhancing the network's learning of feature information and improving the quality of CAM. Moreover, to further improve the segmentation precision, we design a Pseudo Label Retraining Strategy (PLRS). This strategy refines the accuracy in local regions, elevates the quality of pseudo labels, and aims to enhance segmentation precision. Experimental results on the PASCAL VOC 2012 and MS COCO 2014 datasets show that our method achieves impressive performance among end-to-end WSSS methods.

IS-GNN: Graph neural network enhanced by aggregating influential and structurally similar nodes

Nowadays, it has been demonstrated that graph neural networks (GNNs) are a powerful tool for graph representation learning. However, most GNN models are shallow ones because stacking more layers may cause the over-smoothing problem. In addition, conventional neighborhood aggregation GNNs have an implicit homophily assumption, leading to poor performance when dealing with networks with heterophily. To address the aforementioned issues, we present a new GNN model, called IS-GNN, to effectively enhance the representation quality of nodes in general networks. IS-GNN learns the representation of each node using three aspects of information: local structural information from its neighborhood, global topological information from influential nodes, and homophilous information from its structurally similar nodes. Consequently, the representations of nodes are more powerful and discriminative. The effectiveness of IS-GNN is validated on both homophilous and heterophilous networks. The experimental results of semi-supervised node classification manifest that the accuracy of IS-GNN is superior to that of baselines.

Adaptive Decision Spatio-temporal neural ODE for traffic flow forecasting with Multi-Kernel Temporal Dynamic Dilation Convolution

Traffic flow prediction is crucial for efficient traffic management. It involves predicting vehicle movement patterns to reduce congestion and enhance traffic flow. However, the highly non-linear and complex patterns commonly observed in traffic flow pose significant challenges for this task. Current Graph Neural Network (GNN) models often construct shallow networks, which limits their ability to extract deeper spatio-temporal representations. Neural ordinary differential equations for traffic prediction address over-smoothing but require significant computational resources, leading to inefficiencies, and sometimes deeper networks may lead to poorer predictions for complex traffic information. In this study, we propose an Adaptive Decision spatio-temporal Neural Ordinary Differential Network, which can adaptively determine the number of layers of ODE according to the complexity of traffic information. It can solve the over-smoothing problem better, improving overall efficiency and prediction accuracy. In addition, traditional temporal convolution methods make it difficult to deal with complex and variable traffic time information with a large time span. Therefore, we introduce a multi-kernel temporal dynamic expansive convolution to handle the traffic time information. Multi-kernel temporal dynamic expansive convolution employs a dynamic dilation strategy, dynamically adjusting the network’s receptive field across levels, effectively capturing temporal dependencies, and can better adapt to the changing time data of traffic information. Additionally, multi-kernel temporal dynamic expansive convolution integrates multi-scale convolution kernels, enabling the model to learn features across diverse temporal scales. We evaluated our proposed method on several real-world traffic datasets. Experimental results show that our method outperformed state-of-the-art benchmarks.

A2GCN: Graph Convolutional Networks with Adaptive Frequency and Arbitrary Order

Graph Neural Networks (GNNs) gain remarkable success in various graph learning tasks under homophily graph assumption. This assumption is extremely fragile since real-world graphs with heterophily are ubiquitous. Under this circumstance, existing GNNs attempt to design or learn graph spectral filters for observed graphs. The representation ability of them, however, is limited due to: (1) Constant frequency response is incapable of simulating complicated filters that real-world applications require. (2) Fixed polynomial order fails to effectively uncover the node label patterns concealing different order neighborhoods. To this end, we propose a novel Graph Convolutional Networks with Adaptive Frequency and Arbitrary Order (A2GCN) to learn various graph spectral filters suitable for distinct networks. Specifically, a simple but elegant filter with adaptive frequency response is designed to span across multiple layers for capturing different frequency components hiding in varying orders, producing A2GCN filter bases. Afterward, the coefficients of the A2GCN basis for each node are learned to achieve A2GCN filters with arbitrary polynomial order. The resulting A2GCN filters possess flexible frequency response that can automatically adapt to the node label pattern, as such, it empowers A2GCN with stronger expressiveness and naturally alleviates the over-smoothing problem. Theoretical analysis is provided to show the superiority of the proposed A2GCN. Additionally, extensive experiments on both node-level and graph-level tasks validate that the proposed A2GCN accomplishes highly competitive performance and improves classification accuracy. Codes are available at https://github.com/AIG22/A2GCN.

Multi-Hop Multi-View Memory Transformer for Session-Based Recommendation

A Session-Based Recommendation (SBR) seeks to predict users’ future item preferences by analyzing their interactions with previously clicked items. In recent approaches, Graph Neural Networks (GNNs) have been commonly applied to capture item relations within a session to infer user intentions. However, these GNN-based methods typically struggle with feature ambiguity between the sequential session information and the item conversion within an item graph, which may impede the model’s ability to accurately infer user intentions. In this article, we propose a novel Multi-hop Multi-view Memory Transformer (M 3 T) to effectively integrate the sequence-view information and relation conversion (graph-view information) of items in a session. First, we propose a Multi-view Memory Transformer (M 2 T) module to concurrently obtain multi-view information of items. Then, a set of trainable memory matrices are employed to store sharable item features, which mitigates cross-view item feature ambiguity. To comprehensively capture latent user intentions, an M 3 T framework is designed to integrate user intentions across different hops of an item graph. Specifically, a k-order power method is proposed to manage the item graph to alleviate the over-smoothing problem when obtaining high-order relations of items. Extensive experiments conducted on three real-world datasets demonstrate the superiority of our method.

PatchSkip: A lightweight technique for effectively alleviating over-smoothing in vision transformers

Recently vision transformers (ViTs) have encountered the over-smoothing problem, which reduces their capacity by mapping input patches into a similar latent representation. Existing methods introduce regularization terms to alleviate over-smoothing but often increase computational costs. To address this, this paper proposes PatchSkip, a novel and flexible dropout variant, alleviating the over-smoothing problem of ViTs in a lightweight manner. Specifically, PatchSkip draws inspiration from the fact that a similar over-smoothing problem in GNNs is primarily caused by static adjacent matrices leading to solitary message passing mode between nodes. PatchSkip constructs graphs with patch embeddings and analyzes the adjacent matrices in ViTs. By randomly selecting specific patch embeddings to bypass transformer blocks, PatchSkip is proved to generate various adjacent matrices and acts as a multi-mode message passing engine, providing diverse modes of message passing between patches. The effectiveness of PatchSkip in preventing over-smoothing is demonstrated through theoretical proofs and empirical visualizations. Furthermore, PatchSkip is evaluated on various datasets and backbones, showing significant performance improvements while reducing computational costs. For example, when trained on Tiny-ImageNet from scratch, PatchSkip improves the performance of the vanilla CrossViT by 3.85% while reducing computational costs by over 20%.

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.