Project Articles

The 19th China Conference on Machine Learning (CCML 2023)

Default Latest Most Read

Please wait a minute... For Selected: Download Citations EndNote Ris BibTeX Toggle Thumbnails Select PIPNet： lightweight asphalt pavement crack image segmentation network Jun FENG, Jiankang BI, Yiru HUO, Jiakuan LI Journal of Computer Applications    2024, 44 (5): 1520-1526.   DOI: 10.11772/j.issn.1001-9081.2023050911 Abstract （163）   HTML （14）    PDF （3158KB）（129）       Knowledge map    Save Crack segmentation is an important prerequisite for evaluating the damage degree of pavement diseases. In order to balance the effectiveness and real-time of deep neural network segmentation， a lightweight asphalt pavement crack segmentation neural network based on U?Net encoder-decoder structure was proposed， namely PIPNet （Parallel dilated convolution of Inverted Pyramid Network）. The encoding part was an inverted pyramid structure. Multi-branch parallel dilated convolution module with different dilatation rates was proposed to extract multi-scale information from the top， middle and bottom features and reduce model complexity， which combined deep separable convolutions with ordinary convolutions and gradually reduced the number of parallel convolutions. Drawing on the characteristics of GhostNet， an inverse residual lightweight module was designed， which was embedded with parallel dual pooling attention. Test results on GAPs384 dataset show that， compared with ResNet50 encoding method， PIPNet has mIoU （mean Intersection over Union） 1.10 percentage points higher with only about one-sixth of parameter quantity and MFLOPs （Million FLOating Point operations）， and its mIoU is 4.14 and 9.95 percentage points higher than those of lightweight GhostNet and SegNet， respectively. Experimental results show that PIPNet has high crack segmentation performance while reducing the model complexity， and has good adaptability to segmentation of different pavement crack images. Table and Figures | Reference | Related Articles | Metrics Select Node classification algorithm fusing 2-connected motif-structure information Wenping ZHENG, Huilin GE, Meilin LIU, Gui YANG Journal of Computer Applications    2024, 44 (5): 1464-1470.   DOI: 10.11772/j.issn.1001-9081.2023050846 Abstract （155）   HTML （6）    PDF （1734KB）（143）       Knowledge map    Save Node representation learning has been widely applied in machine learning tasks， such as node classification， clustering and link prediction， since it can encode graph structure data information into low-dimensional potential space. In complex networks， nodes are interacted through not only low-order interactions， but also higher-order interactions formed by special connection modes. The higher-order interactions of a complex network are also called motifs. A node classification algorithm Fusing 2-connected Motif-structure Information （FMI） was proposed to use motif information among nodes to obtain node representation for node classification tasks. Firstly， the 2-connected motifs in the network were counted. A measure index of node importance， named motif-ratio， was proposed by using the motif information in the node； and a sampling probability was calculated according to the motif-ratio to carry out neighborhood sampling. A weighted auxiliary graph was constructed to fuse the low-order relations and the high-order relations of network nodes to aggregate neighborhoods weightedly. The node classification was performed on 5 datasets， Cora， Citeseer， Pubmed， Wiki and DBLP. By comparing with 5 classical baseline algorithms， the proposed algorithm FMI shows better performance on Accuracy， F1-score and other indicators. Table and Figures | Reference | Related Articles | Metrics Select Robust learning method by reweighting examples with negative learning Boshi ZOU, Ming YANG, Chenchen ZONG, Mingkun XIE, Shengjun HUANG Journal of Computer Applications    2024, 44 (5): 1479-1484.   DOI: 10.11772/j.issn.1001-9081.2023050880 Abstract （237）   HTML （5）    PDF （1241KB）（328）       Knowledge map    Save Noisy label learning methods can effectively use data containing noisy labels to train models and significantly reduce the labeling cost of large-scale datasets. Most existing noisy label learning methods usually assume that the number of each class in the dataset is balanced， but the data in many real-world scenarios tend to have noisy labels， and long-tailed distributions often present in the dataset simultaneously， making it difficult for existing methods to select clean examples from noisy examples in the tail class according to traning loss or confidence. To solve noisy long-tailed learning problem， a ReWeighting examples with Negative Learning （NLRW） method was proposed， by which examples were reweighted adaptively based on negative learning. Specifically， at each training epoch， the weights of examples were calculated according to the output distributions of the model to head classes and tail classes. The weights of clean examples were close to one while the weights of noisy examples were close to zero. To ensure accurate estimation of weights， negative learning and cross entropy loss were combined to train the model with a weighted loss function. Experimental results on CIFAR-10 and CIFAR-100 datasets with various imbalance rates and noise rates show that， compared with the optimal baseline model TBSS （Two stage Bi-dimensional Sample Selection） for noisy long-tail classification， NLRW method improves the average accuracy by 4.79% and 3.46%， respectively. Table and Figures | Reference | Related Articles | Metrics Select Point cloud classification network based on node structure Wenshuo GAO, Xiaoyun CHEN Journal of Computer Applications    2024, 44 (5): 1471-1478.   DOI: 10.11772/j.issn.1001-9081.2023050802 Abstract （215）   HTML （16）    PDF （2562KB）（420）       Knowledge map    Save The non-structured and non-uniform distribution of point cloud data poses significant challenges for feature representation and classification tasks. To extract the three-dimensional structural features of point cloud objects， existing methods often employ complex local feature extraction structures to construct hierarchical networks， resulting in a complex feature extraction network that mainly focuses on the local structures of the point cloud objects. To better extract features from unevenly distributed point cloud objects， a Node structure Network （NsNet） with sample point convolution density adaptive weighting was proposed. The convolutional network adaptively weighted sample points based on Gaussian density to differentiate the density differences among sampling points， thereby better characterizing the overall structure of objects. Additionally， the network structure was simplified by incorporating spherical coordinates to reduce model complexity. Experimental results on three public datasets demonstrate that， NsNet based on adaptive density weighting improves the Overall Accuracy （OA） by 9.1 and 1.3 percentage points respectively compared with PointNet++ and PointMLP， and reduces the number of parameters by 4.6×106 compared to PointMLP. NsNet can effectively address the problem of information loss caused by uneven distribution of point clouds， improve the classification accuracy and reduce the model complexity. Table and Figures | Reference | Related Articles | Metrics Select Task offloading method based on dynamic service cache assistance Junna ZHANG, Xinxin WANG, Tianze LI, Xiaoyan ZHAO, Peiyan YUAN Journal of Computer Applications    2024, 44 (5): 1493-1500.   DOI: 10.11772/j.issn.1001-9081.2023050831 Abstract （205）   HTML （6）    PDF （2414KB）（327）       Knowledge map    Save Aiming at the problem of user experience quality degradation due to the lack of comprehensive consideration of the diversity and dynamics of user service requests in the joint optimization of service caching and task offloading， a task offloading method based on dynamic service cache assistance was proposed. Firstly， to address the problem of the large action spaces for edge servers performing caching service， the actions were redefined and the optimal set of actions was selected to improve the efficiency of algorithm training. Secondly， an improved multi-agent Q-Learning algorithm was designed to learn an optimal service caching policy. Thirdly， the task offloading problem was converted into a convex optimization problem， and the optimal solution was obtained using a convex optimization tool. Finally， the optimal computational resource allocation policy was found using the Lagrangian dual method. To verify the effectiveness of the proposed method， extensive experiments were conducted based on a real dataset. Experimental results show that the response time of the proposed method is reduced by 8.5%， 11.8% and 12.6%， respectively， and the average quality of experience is improved by 1.5%， 2.7% and 4.3%， respectively， compared with Q-Learning， Double Deep Q Network （D2QN） and Multi-Agent Deep Deterministic Policy Gradient （MADDPG） method. Table and Figures | Reference | Related Articles | Metrics Select Collaborative offloading strategy in internet of vehicles based on asynchronous deep reinforcement learning Xiaoyan ZHAO, Wei HAN, Junna ZHANG, Peiyan YUAN Journal of Computer Applications    2024, 44 (5): 1501-1510.   DOI: 10.11772/j.issn.1001-9081.2023050788 Abstract （251）   HTML （9）    PDF （2661KB）（352）       Knowledge map    Save With the rapid development of Internet of Vehicles （IoV）， smart connected vehicles generate a large number of latency-sensitive and computation-intensive tasks， and limited vehicle computing resources and traditional cloud service modes cannot meet the needs of in-vehicle users. Mobile Edge Computing （MEC） provides an effective paradigm for solving task offloading of massive data. However， when considering multi-task and multi-user scenarios， the complexity of task offloading scenarios in IoV is high due to the real-time and dynamic changes in vehicle locations， task types and vehicle density， and the offloading process is prone to problems such as unbalanced edge resource allocation， excessive communication cost overhead and slow algorithm convergence. To solve the above problems， cooperative task offloading strategy of multiple edge servers in multi-task and multi-user mobile scenarios in IoV was focused on. First， a three-layer heterogeneous network model for multi-edge collaborative processing was proposed， and dynamic collaborative clusters were introduced for the changing environment in IoV to transform the offloading problem into a joint optimization problem of delay and energy consumption. Then， the problem was divided into two subproblems of offloading decision and resource allocation， where the resource allocation problem was further split into resource allocation for edge servers and transmission bandwidth， and the two subproblems were solved based on convex optimization theory. In order to find the optimal offloading decision set， a Multi-edge Collaborative Deep Deterministic Policy Gradient （MC-DDPG） algorithm that can handle continuous problems in collaborative clusters was proposed， based on which an Asynchronous MC-DDPG （AMC-DDPG） algorithm was designed. The training parameters in collaborative clusters were asynchronously uploaded to the cloud for global update， and then the updated results were returned to each collaborative cluster to improve the convergence speed. Simulation results show that the AMC-DDPG algorithm improves the convergence speed by at least 30% over the DDPG algorithm and achieves better results in terms of reward and total cost. Table and Figures | Reference | Related Articles | Metrics Select PAGCL： positive augmentation graph contrastive learning recommendation method without negative sampling Jiong WANG, Taotao TANG, Caiyan JIA Journal of Computer Applications    2024, 44 (5): 1485-1492.   DOI: 10.11772/j.issn.1001-9081.2023050756 Abstract （291）   HTML （19）    PDF （2404KB）（693）       Knowledge map    Save Contrastive Learning （CL） has been widely used for recommendation because of its ability to extract supervised signals contained in data itself. The recent study shows that the success of CL in recommendation depends on the uniformity of node distribution brought by comparative loss — Infomation Noise Contrastive Estimation （InfoNCE） loss. In addition， the other study proves that Bayesian Personalized Ranking （BPR） loss is beneficial to alignment and uniformity， which contribute to higher recommendation performance. Since the CL loss can bring stronger uniformity than the negative term of BPR， the necessity of the negative term of BPR in CL framework has aroused suspicion. Therefore， this study experimentally disclosed that the negative term of BPR is unnecessary in CL framework for recommendation. Based on this observation， a joint optimization loss without negative sampling was proposed， which could be applied to classical CL-based methods and achieve the same or higher performance. Besides， unlike studies which focus on improving uniformity， a novel Positive Augmentation Graph Contrastive Learning method （PAGCL） was presented， which used random positive samples for perturbation at representation level to further strengthen alignment. Experimental results on several benchmark datasets show that the proposed method is superior to SOTA （State-Of-The-Art） methods like Self-supervised Graph Learning （SGL） and Simple Graph Contrastive Learning （SimGCL） on recall and Normalized Discounted Cumulative Gain （NDCG）. The method’s improvement over the base model Light Graph Convolutional Network （LightGCN） can reach up to 17.6% at NDCG@20. Table and Figures | Reference | Related Articles | Metrics Select Driver behavior recognition based on dual-path spatiotemporal network Zhiyuan XI, Chao TANG, Anyang TONG, Wenjian WANG Journal of Computer Applications    2024, 44 (5): 1511-1519.   DOI: 10.11772/j.issn.1001-9081.2023050800 Abstract （246）   HTML （14）    PDF （3642KB）（231）       Knowledge map    Save Dangerous driving behavior of drivers is one of the main causes of vicious traffic accidents， so identifying driver’s behavior is of great significance for engineering applications. Currently， the mainstream vision-based detection methods are to study the local spatiotemporal features of driver behavior， and less research is done on global spatial features and long-term temporal correlation features， which to a certain extent cannot be combined with the scene context information to identify dangerous driving behaviors. To solve the above problems， a driver behavior recognition method based on a dual-path spatiotemporal network was proposed， which integrated the advantages of different spatiotemporal pathways to improve the richness of behavioral features. Firstly， an improved Two-Stream convolutional Network （TSN） was used to learn the spatiotemporal information for characterization while reducing the sparsity of extracted features. Secondly， a Transformer-based serial spatiotemporal network was constructed to supplement the long-term temporal correlation information. Finally， a fusion decision was made using a dual-path spatiotemporal network to enhance the robustness of the model. Experimental results show that the proposed method achieves recognition accuracies of 99.85%， 99.94% and 98.77% on three publicly available datasets： a driver fatigue detection dataset YawDD， a driver distraction detection dataset SF-DDDD （State-Farm Distracted Driver Detection Dataset）， and a the latest driver behavior recognition dataset SynDD1， respectively； especially on SynDD1， the recognition accuracy is improved by 1.64 percentage points compared to MoviNet-A0， a recognition network by motion. Ablation experimental results confirm that the proposed method has high recognition accuracy of driver behavior. Table and Figures | Reference | Related Articles | Metrics