Due to the conditional independent assumption of a CTC model, a language model is usually added to improve its speech recognition performance. However, adding a language model will increase the complexity and computation cost. Therefore, we proposed a simple and effective speech recognition method based on CTC multilayer loss. Unlike the traditional CTC model which only optimizes the CTC loss of the last layer, in this method, the CTC multilayer loss, which guides the training of the model, is obtained by weighted summation of the CTC losses of different layers. Through optimizing the losses of different layers, the information of different layers of the CTC model can be taken into account, and the information obtained is more comprehensive, so that the model obtained has better recognition performance. With a small amount of code modification, this CTC multilayer loss method can well regulate the training of CTC and improve the performance of speech recognition. Since this method only changes the loss function of the CTC model and does not change the structure of the CTC model and its testing process, the training stage is simple and the testing stage has no extra memory cost and computation cost. We evaluated the method on Aishell-1 dataset using WeNet as the baseline, and it was able to reduce the character error rate (CER) by 7.5% and improve speech recognition performance without adding a language model.
{"title":"Speech Recognition Method based on CTC Multilayer Loss","authors":"Deyu Luo, Xianhong Chen, Mao-shen Jia, C. Bao","doi":"10.1145/3581807.3581864","DOIUrl":"https://doi.org/10.1145/3581807.3581864","url":null,"abstract":"Due to the conditional independent assumption of a CTC model, a language model is usually added to improve its speech recognition performance. However, adding a language model will increase the complexity and computation cost. Therefore, we proposed a simple and effective speech recognition method based on CTC multilayer loss. Unlike the traditional CTC model which only optimizes the CTC loss of the last layer, in this method, the CTC multilayer loss, which guides the training of the model, is obtained by weighted summation of the CTC losses of different layers. Through optimizing the losses of different layers, the information of different layers of the CTC model can be taken into account, and the information obtained is more comprehensive, so that the model obtained has better recognition performance. With a small amount of code modification, this CTC multilayer loss method can well regulate the training of CTC and improve the performance of speech recognition. Since this method only changes the loss function of the CTC model and does not change the structure of the CTC model and its testing process, the training stage is simple and the testing stage has no extra memory cost and computation cost. We evaluated the method on Aishell-1 dataset using WeNet as the baseline, and it was able to reduce the character error rate (CER) by 7.5% and improve speech recognition performance without adding a language model.","PeriodicalId":292813,"journal":{"name":"Proceedings of the 2022 11th International Conference on Computing and Pattern Recognition","volume":"13 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129098358","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Developing new drugs is time-consuming, labor-intensive and costly. Identifying new targets for existing drugs can help to discover new potential therapeutic uses of old drugs and reduce the cost of drug development. Drug-target interactions are usually inferred by searching for similar drugs and targets. Various biomedical databases have been established currently, which provide effective data for predicting drug-target interactions. We proposed a novel computational model for discovering Drug-Target Interactions using Network consistency project (DTIN). The Gaussian kernel similarity of drugs and targets were derived from known drug-target interactions by Gaussian kernel function, thus DTIN incorporated six types of similarities, including drug chemical structure similarity, drug ATC similarity, drug Gaussian kernel similarity, target sequence similarity, target function similarity, and target Gaussian kernel similarity. We used logistic regression to process the integrated similarity and predicted scores of interacting drug-target pairs by network consistency projection. Five-fold cross-validation was implemented on a benchmark dataset, and the computational results demonstrated that DTIN was effective and outperformed two advanced models.
{"title":"Inferring Potential Drug-Target Interactions using Multiple Similarities and Network Consistency Projection","authors":"Jianhua Li, Haoran Ren, Dayu Xiao, Botao Deng","doi":"10.1145/3581807.3581860","DOIUrl":"https://doi.org/10.1145/3581807.3581860","url":null,"abstract":"Developing new drugs is time-consuming, labor-intensive and costly. Identifying new targets for existing drugs can help to discover new potential therapeutic uses of old drugs and reduce the cost of drug development. Drug-target interactions are usually inferred by searching for similar drugs and targets. Various biomedical databases have been established currently, which provide effective data for predicting drug-target interactions. We proposed a novel computational model for discovering Drug-Target Interactions using Network consistency project (DTIN). The Gaussian kernel similarity of drugs and targets were derived from known drug-target interactions by Gaussian kernel function, thus DTIN incorporated six types of similarities, including drug chemical structure similarity, drug ATC similarity, drug Gaussian kernel similarity, target sequence similarity, target function similarity, and target Gaussian kernel similarity. We used logistic regression to process the integrated similarity and predicted scores of interacting drug-target pairs by network consistency projection. Five-fold cross-validation was implemented on a benchmark dataset, and the computational results demonstrated that DTIN was effective and outperformed two advanced models.","PeriodicalId":292813,"journal":{"name":"Proceedings of the 2022 11th International Conference on Computing and Pattern Recognition","volume":"45 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129104145","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Medium- and long-wave infrared image fusion has problems such as overemphasizing detail retention, which often weakens the presence of thermal information, poor contrast of fused images, and large noise, so a medium- and long-wave image fusion method based on improved non-subsample shearlt transform (NSST) is proposed. Firstly, the image processing of mid-wave infrared and long-wave infrared images is carried out in a targeted manner, and the pixel values of the target and background area are adjusted by using the adaptive contrast enhancement algorithm to adjust the pixel values of the mid-wave infrared image, so as to achieve the target enhancement effect by expanding the relative pixel difference between the thermal target and the background area. Secondly, the average curvature filtering and Gaussian filtering are used to decompose the source image into detail layer, structure layer and area layer. The energy differential feature is used to guide the energy attribute fusion strategy to fuse the regional layer, the structure layer adopts the maximum fusion strategy to fuse, and the detail layer adopts the fusion strategy of directional contrast. Finally, the three levels after fusion are added to reconstruct the final fusion image. Experimental results show that the algorithm can effectively fuse mid-wave infrared and long-wave infrared images, which can not only effectively retain the mid-wave infrared thermal radiation and heat information, but also retain the edge detail expression ability in the fusion results to a large extent. It can be seen from the subjective and objective evaluation indicators that the proposed algorithm shows better fusion performance than other algorithms.
{"title":"Medium and Long Wave Infrared Image Enhancement Fusion Method Based on Edge Preserving","authors":"Shubin Lou, Xin Zheng, Bin Yue, Qiang Wu","doi":"10.1145/3581807.3581852","DOIUrl":"https://doi.org/10.1145/3581807.3581852","url":null,"abstract":"Medium- and long-wave infrared image fusion has problems such as overemphasizing detail retention, which often weakens the presence of thermal information, poor contrast of fused images, and large noise, so a medium- and long-wave image fusion method based on improved non-subsample shearlt transform (NSST) is proposed. Firstly, the image processing of mid-wave infrared and long-wave infrared images is carried out in a targeted manner, and the pixel values of the target and background area are adjusted by using the adaptive contrast enhancement algorithm to adjust the pixel values of the mid-wave infrared image, so as to achieve the target enhancement effect by expanding the relative pixel difference between the thermal target and the background area. Secondly, the average curvature filtering and Gaussian filtering are used to decompose the source image into detail layer, structure layer and area layer. The energy differential feature is used to guide the energy attribute fusion strategy to fuse the regional layer, the structure layer adopts the maximum fusion strategy to fuse, and the detail layer adopts the fusion strategy of directional contrast. Finally, the three levels after fusion are added to reconstruct the final fusion image. Experimental results show that the algorithm can effectively fuse mid-wave infrared and long-wave infrared images, which can not only effectively retain the mid-wave infrared thermal radiation and heat information, but also retain the edge detail expression ability in the fusion results to a large extent. It can be seen from the subjective and objective evaluation indicators that the proposed algorithm shows better fusion performance than other algorithms.","PeriodicalId":292813,"journal":{"name":"Proceedings of the 2022 11th International Conference on Computing and Pattern Recognition","volume":"269 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122471767","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Handwritten mathematical expression recognition (HMER) is a challenging task due to the complex two-dimensional structure of mathematical expressions and the similarity of handwritten texts. Most existing methods for HMER only consider single-scale features while ignoring multi-scale features that are very important to HMER. Few works have explored the fusion of multi-scale features in HMER, but exhibited an extra branch that brings more parameters and computation. In this paper, we propose an end-to-end method to integrate multi-scale features using a unified model. Specifically, we customized the Dense Atrous Spatial Pyramid Pooling (DenseASPP) to our backbone network to capture the multi-scale features of the input image meanwhile expanding the receptive fields. Moreover, we added a symbol classifier using focal loss to better discriminate and recognize similar symbols, to further improve the performance of HMER. Experiments on the Competition on Recognition of Online Handwritten Mathematical Expressions (CROHME) 2014, 2016 and 2019 shows that the proposed method achieves superior performance to most state-of-the-art methods, demonstrating the effectiveness of the proposed method.
{"title":"Multi-Scale Features Integration for Handwritten Mathematical Expression Recognition","authors":"Xianghao Liu, Da-han Wang, Shunzhi Zhu","doi":"10.1145/3581807.3581844","DOIUrl":"https://doi.org/10.1145/3581807.3581844","url":null,"abstract":"Handwritten mathematical expression recognition (HMER) is a challenging task due to the complex two-dimensional structure of mathematical expressions and the similarity of handwritten texts. Most existing methods for HMER only consider single-scale features while ignoring multi-scale features that are very important to HMER. Few works have explored the fusion of multi-scale features in HMER, but exhibited an extra branch that brings more parameters and computation. In this paper, we propose an end-to-end method to integrate multi-scale features using a unified model. Specifically, we customized the Dense Atrous Spatial Pyramid Pooling (DenseASPP) to our backbone network to capture the multi-scale features of the input image meanwhile expanding the receptive fields. Moreover, we added a symbol classifier using focal loss to better discriminate and recognize similar symbols, to further improve the performance of HMER. Experiments on the Competition on Recognition of Online Handwritten Mathematical Expressions (CROHME) 2014, 2016 and 2019 shows that the proposed method achieves superior performance to most state-of-the-art methods, demonstrating the effectiveness of the proposed method.","PeriodicalId":292813,"journal":{"name":"Proceedings of the 2022 11th International Conference on Computing and Pattern Recognition","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115773702","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Shuai Hou, Jizhe Lu, Enguo Zhu, Hailong Zhang, Aliaosha Ye
To improve the efficiency of data collection, transmission and application in the electric power Internet of Things(IoT), many researches are devoted to resource allocation and scheduling algorithms. However, few studies focus on the impact of dynamic changes in data volume on decision-making. In this paper, we propose an intelligent IoT scheduling mechanism based on data traffic prediction. First, we propose an IoT data traffic prediction model(IoT-DTP) to accurately predict the future data volume. On this basis, we construct a data-driven IoT scheduling mechanism (PESM), which can realize higher real-time data transmission capability and faster service response. For instance, it can realize efficient data interaction of App launch, release and update in the intelligent IoT software platform. Finally, through theoretical analysis and experimental evaluation, the efficiency of the proposed method is verified.
{"title":"Intelligent IoT Scheduling Mechanism Based on Data Traffic Prediction","authors":"Shuai Hou, Jizhe Lu, Enguo Zhu, Hailong Zhang, Aliaosha Ye","doi":"10.1145/3581807.3581899","DOIUrl":"https://doi.org/10.1145/3581807.3581899","url":null,"abstract":"To improve the efficiency of data collection, transmission and application in the electric power Internet of Things(IoT), many researches are devoted to resource allocation and scheduling algorithms. However, few studies focus on the impact of dynamic changes in data volume on decision-making. In this paper, we propose an intelligent IoT scheduling mechanism based on data traffic prediction. First, we propose an IoT data traffic prediction model(IoT-DTP) to accurately predict the future data volume. On this basis, we construct a data-driven IoT scheduling mechanism (PESM), which can realize higher real-time data transmission capability and faster service response. For instance, it can realize efficient data interaction of App launch, release and update in the intelligent IoT software platform. Finally, through theoretical analysis and experimental evaluation, the efficiency of the proposed method is verified.","PeriodicalId":292813,"journal":{"name":"Proceedings of the 2022 11th International Conference on Computing and Pattern Recognition","volume":"34 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"132514585","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Vehicle re-identification (Re-ID) aims to retrieve the target vehicle from a large dataset composed of vehicle images captured by multiple cameras. Most vehicles are difficult to recognize in the environment of low resolution, occlusion, and viewpoint change, which brings challenges to vehicle Re-ID. Existing work usually uses additional attribute information to distinguish different vehicles, such as color, viewpoint, and model. However, this requires expensive manual annotation. Therefore, we propose a three-branch network based on attention mechanism and local-global feature association (AM-LGFA) to improve the accuracy of vehicle Re-ID. In the global branch, the global features of the vehicle are extracted. A multi-scale channel attention module is introduced into the attention branch to suppress irrelevant information and extract important channel features. The features extracted from the backbone are divided into different stripe features in the horizontal direction in the local branch. Then connect each stripe feature with the global information to enhance the context between features. Finally, the features extracted from the three branches are concatenated as the feature representation of the test phase. The experimental results show that the features extracted by the AM-LGFA network are complementary. The effectiveness of this method is verified on two challenging public datasets, VehicleID and VeRi-776.
{"title":"Combining Attention Mechanism and Local-Global Features Association Network for Vehicle Re-identification","authors":"Caiyu Li, X. Du, Yun Wu, Da-han Wang","doi":"10.1145/3581807.3581842","DOIUrl":"https://doi.org/10.1145/3581807.3581842","url":null,"abstract":"Vehicle re-identification (Re-ID) aims to retrieve the target vehicle from a large dataset composed of vehicle images captured by multiple cameras. Most vehicles are difficult to recognize in the environment of low resolution, occlusion, and viewpoint change, which brings challenges to vehicle Re-ID. Existing work usually uses additional attribute information to distinguish different vehicles, such as color, viewpoint, and model. However, this requires expensive manual annotation. Therefore, we propose a three-branch network based on attention mechanism and local-global feature association (AM-LGFA) to improve the accuracy of vehicle Re-ID. In the global branch, the global features of the vehicle are extracted. A multi-scale channel attention module is introduced into the attention branch to suppress irrelevant information and extract important channel features. The features extracted from the backbone are divided into different stripe features in the horizontal direction in the local branch. Then connect each stripe feature with the global information to enhance the context between features. Finally, the features extracted from the three branches are concatenated as the feature representation of the test phase. The experimental results show that the features extracted by the AM-LGFA network are complementary. The effectiveness of this method is verified on two challenging public datasets, VehicleID and VeRi-776.","PeriodicalId":292813,"journal":{"name":"Proceedings of the 2022 11th International Conference on Computing and Pattern Recognition","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130888493","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
This paper designs a multi-branch feature fusion classification algorithm to improve the network accuracy of the classical deep learning-based infrared image algorithms. First, the algorithm uses a multi-resolution sub-network parallel connection method to build the overall network architecture. Then, a lightweight structural module is designed to reduce the computational load of network weight parameters, and a channel attention module is introduced to refine feature channels and improve detection accuracy. Finally, the parallel connection mode of the spatial pyramid is designed to enhance the ability of feature semantic expression. The experimental results show the improved accuracy of the algorithm model proposed in this paper and the optimization of parameters. The accuracy rate can reach 97.6%. The proposed algorithm is an innovation to the current mainstream classification algorithm, which reflects good promotion and application.
{"title":"PV infrared hot spot classification algorithm with multi-branch feature fusion","authors":"Han Zhou","doi":"10.1145/3581807.3581836","DOIUrl":"https://doi.org/10.1145/3581807.3581836","url":null,"abstract":"This paper designs a multi-branch feature fusion classification algorithm to improve the network accuracy of the classical deep learning-based infrared image algorithms. First, the algorithm uses a multi-resolution sub-network parallel connection method to build the overall network architecture. Then, a lightweight structural module is designed to reduce the computational load of network weight parameters, and a channel attention module is introduced to refine feature channels and improve detection accuracy. Finally, the parallel connection mode of the spatial pyramid is designed to enhance the ability of feature semantic expression. The experimental results show the improved accuracy of the algorithm model proposed in this paper and the optimization of parameters. The accuracy rate can reach 97.6%. The proposed algorithm is an innovation to the current mainstream classification algorithm, which reflects good promotion and application.","PeriodicalId":292813,"journal":{"name":"Proceedings of the 2022 11th International Conference on Computing and Pattern Recognition","volume":"11 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130656034","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Aiming at the severe form of new coronavirus epidemic prevention and control, a target detection algorithm is proposed to detect whether masks are worn in public places. The Ghostnet and SElayer modules with fewer design parameters replace the BottleneckCSP part in the original Yolov5s network, which reduces the computational complexity of the model and improves the detection accuracy. The bounding box regression loss function DIOU is optimized, the DGIOU loss function is used for bounding box regression, and the center coordinate distance between the two bounding boxes is considered to achieve a better convergence effect. In the feature pyramid, the depthwise separable convolution DW is used to replace the ordinary convolution, which further reduces the amount of parameters and reduces the loss of feature information caused by multiple convolutions. The experimental results show that compared with the yolov5s algorithm, the proposed method improves the mAP by 4.6% and the detection rate by 10.7 frame/s in the mask wearing detection. Compared with other mainstream algorithms, the improved yolov5s algorithm has better generalization ability and practicability.
{"title":"A Mask Detection Algorithm Based on Improved Yolov5s","authors":"Xin Zhang, Yalan Zeng, Shunyong Zhou","doi":"10.1145/3581807.3581818","DOIUrl":"https://doi.org/10.1145/3581807.3581818","url":null,"abstract":"Aiming at the severe form of new coronavirus epidemic prevention and control, a target detection algorithm is proposed to detect whether masks are worn in public places. The Ghostnet and SElayer modules with fewer design parameters replace the BottleneckCSP part in the original Yolov5s network, which reduces the computational complexity of the model and improves the detection accuracy. The bounding box regression loss function DIOU is optimized, the DGIOU loss function is used for bounding box regression, and the center coordinate distance between the two bounding boxes is considered to achieve a better convergence effect. In the feature pyramid, the depthwise separable convolution DW is used to replace the ordinary convolution, which further reduces the amount of parameters and reduces the loss of feature information caused by multiple convolutions. The experimental results show that compared with the yolov5s algorithm, the proposed method improves the mAP by 4.6% and the detection rate by 10.7 frame/s in the mask wearing detection. Compared with other mainstream algorithms, the improved yolov5s algorithm has better generalization ability and practicability.","PeriodicalId":292813,"journal":{"name":"Proceedings of the 2022 11th International Conference on Computing and Pattern Recognition","volume":"70 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125349361","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Kemeng Wang, Quan Zhou, Zhikang Zeng, Menglong Chen
With the emergence of encrypted traffic, more and more researchers use AI technology to improve the accuracy of traffic identification. However, machine learning needs to rely on human experience to extract features, and the training of deep learning models depends on a large number of labeled samples.To solve these problems, we propose an encrypted traffic identification method based on RepVGG. First, the pre-trained model RepVGG-A0 on the ImageNet dataset is migrated to the encrypted traffic dataset, and a dropout layer is added before the linear classifier in order to avoid overfitting. Then, to reduce the impact of sample imbalance, different weight parameters are assigned to different categories in the training process.Finally, we make a comparison with other traffic identification methods.The experimental results show that the proposed method can achieve 99.98% accuracy in binary classification and 97% accuracy in multi-classification experiments, which proves the effectiveness of the method.
{"title":"An Encrypted Traffic Identification Method Based on RepVGG","authors":"Kemeng Wang, Quan Zhou, Zhikang Zeng, Menglong Chen","doi":"10.1145/3581807.3581896","DOIUrl":"https://doi.org/10.1145/3581807.3581896","url":null,"abstract":"With the emergence of encrypted traffic, more and more researchers use AI technology to improve the accuracy of traffic identification. However, machine learning needs to rely on human experience to extract features, and the training of deep learning models depends on a large number of labeled samples.To solve these problems, we propose an encrypted traffic identification method based on RepVGG. First, the pre-trained model RepVGG-A0 on the ImageNet dataset is migrated to the encrypted traffic dataset, and a dropout layer is added before the linear classifier in order to avoid overfitting. Then, to reduce the impact of sample imbalance, different weight parameters are assigned to different categories in the training process.Finally, we make a comparison with other traffic identification methods.The experimental results show that the proposed method can achieve 99.98% accuracy in binary classification and 97% accuracy in multi-classification experiments, which proves the effectiveness of the method.","PeriodicalId":292813,"journal":{"name":"Proceedings of the 2022 11th International Conference on Computing and Pattern Recognition","volume":"6 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126656011","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Collaborative filtering algorithms have serious data sparsity and cold start problems as the amount of data increases and the movie dataset keeps growing.To solve the above problems, this paper proposes to combine the knowledge graph with Matrix factorization algorithm.Through the user's historical interests, mining the user's similar interests on the knowledge graph, to form the candidate items, useing eventually to predict users' interests, and finally using Bayesian personalized recommendation to predict the user's rating of the candidate items to achieve top-K recommendation.Through experiments, it is demonstrated that the algorithm proposed in this paper significantly improves the recommendation effect of matrix decomposition model. With its AUC=0.9348 and ACC=0.8474 on the movie dataset, the experimental data show that the algorithm can improve the recommendation effect more effectively.
{"title":"Bayesian Personalized Ranking based on Knowledge Graph","authors":"Ran Ma, Xiaotian Yang, Jiang Li, Fei Gao","doi":"10.1145/3581807.3581887","DOIUrl":"https://doi.org/10.1145/3581807.3581887","url":null,"abstract":"Collaborative filtering algorithms have serious data sparsity and cold start problems as the amount of data increases and the movie dataset keeps growing.To solve the above problems, this paper proposes to combine the knowledge graph with Matrix factorization algorithm.Through the user's historical interests, mining the user's similar interests on the knowledge graph, to form the candidate items, useing eventually to predict users' interests, and finally using Bayesian personalized recommendation to predict the user's rating of the candidate items to achieve top-K recommendation.Through experiments, it is demonstrated that the algorithm proposed in this paper significantly improves the recommendation effect of matrix decomposition model. With its AUC=0.9348 and ACC=0.8474 on the movie dataset, the experimental data show that the algorithm can improve the recommendation effect more effectively.","PeriodicalId":292813,"journal":{"name":"Proceedings of the 2022 11th International Conference on Computing and Pattern Recognition","volume":"13 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"114501493","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：