Knowledge-Based Systems最新文献_第3页

Long-term traffic speed prediction utilizing data augmentation via segmented time frame clustering 通过分段时间框架聚类利用数据增强进行长期交通速度预测

IF 7.2 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Knowledge-Based Systems

Pub Date : 2024-11-22 DOI: 10.1016/j.knosys.2024.112785

Robin Kuok Cheong Chan , Joanne Mun-Yee Lim , Rajendran Parthiban

Among many traffic forecasting studies, comparatively fewer studies focus on long-term traffic prediction, such as 24-hour prediction. While traffic data such as traffic speed are easier to obtain, obtaining similarly reliable and accessible feature data with the inclusion of weather or events would be difficult depending on the location or availability of the service providers. Getting these data becomes a more significant issue when considering global coverage. To mitigate the issue of limited feature data, a method to augment already existing data by improving the dataset's quality and ensuring more accurate training via sorting the dataset into appropriate clusters to be used as an additional feature is proposed. This paper proposes a long-term traffic forecasting model that utilizes a novel time-series segmentation method paired with data clustering and classification via Convolutional Neural Network (CNN) to cover the lack of traffic data and features as additional pre-processing before using Long Short-Term Memory (LSTM) for long-term traffic prediction which is not researched as much. This proposed model is called Cluster Augmented LSTM (CAL). The proposed model is compared with existing machine learning models and evaluated using Mean Absolute Percentage Error (MAPE) and Root-Mean-Squared-Error (RMSE) performance metrics. A comparison between LSTM and Gated Recurrent Units (GRU) was conducted, showing that GRU tends to outperform LSTM in most cases. However, the best-performing result for the proposed method still utilizes LSTM. The final results show that the proposed CAL model could achieve better results by 1.42 %-1.76 % and 0.25–0.41 for MAPE and RMSE, respectively.

在众多交通预测研究中，侧重于长期交通预测（如 24 小时预测）的研究相对较少。虽然交通速度等交通数据比较容易获得，但根据服务提供商的地理位置或可用性，要获得包含天气或事件的类似可靠且可访问的特征数据则非常困难。当考虑到全球覆盖范围时，获取这些数据就成了一个更重要的问题。为了缓解特征数据有限的问题，本文提出了一种方法，通过将数据集分类为适当的群组作为附加特征来提高数据集的质量并确保更准确的训练，从而增强现有数据。本文提出了一种长期交通预测模型，利用一种新颖的时间序列分割方法，并通过卷积神经网络（CNN）进行数据聚类和分类，以弥补交通数据和特征的不足，作为使用长短期记忆（LSTM）进行长期交通预测之前的额外预处理。该模型被称为集群增强 LSTM（CAL）。该模型与现有的机器学习模型进行了比较，并使用平均绝对百分比误差 (MAPE) 和均方根误差 (RMSE) 性能指标进行了评估。对 LSTM 和门控递归单元（GRU）进行了比较，结果表明 GRU 在大多数情况下往往优于 LSTM。不过，所提方法中表现最好的仍然是 LSTM。最终结果表明，拟议的 CAL 模型可以取得更好的结果，MAPE 和 RMSE 分别为 1.42 %-1.76 % 和 0.25-0.41 %。

{"title":"Long-term traffic speed prediction utilizing data augmentation via segmented time frame clustering","authors":"Robin Kuok Cheong Chan , Joanne Mun-Yee Lim , Rajendran Parthiban","doi":"10.1016/j.knosys.2024.112785","DOIUrl":"10.1016/j.knosys.2024.112785","url":null,"abstract":"<div><div>Among many traffic forecasting studies, comparatively fewer studies focus on long-term traffic prediction, such as 24-hour prediction. While traffic data such as traffic speed are easier to obtain, obtaining similarly reliable and accessible feature data with the inclusion of weather or events would be difficult depending on the location or availability of the service providers. Getting these data becomes a more significant issue when considering global coverage. To mitigate the issue of limited feature data, a method to augment already existing data by improving the dataset's quality and ensuring more accurate training via sorting the dataset into appropriate clusters to be used as an additional feature is proposed. This paper proposes a long-term traffic forecasting model that utilizes a novel time-series segmentation method paired with data clustering and classification via Convolutional Neural Network (CNN) to cover the lack of traffic data and features as additional pre-processing before using Long Short-Term Memory (LSTM) for long-term traffic prediction which is not researched as much. This proposed model is called Cluster Augmented LSTM (CAL). The proposed model is compared with existing machine learning models and evaluated using Mean Absolute Percentage Error (MAPE) and Root-Mean-Squared-Error (RMSE) performance metrics. A comparison between LSTM and Gated Recurrent Units (GRU) was conducted, showing that GRU tends to outperform LSTM in most cases. However, the best-performing result for the proposed method still utilizes LSTM. The final results show that the proposed CAL model could achieve better results by 1.42 %-1.76 % and 0.25–0.41 for MAPE and RMSE, respectively.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"308 ","pages":"Article 112785"},"PeriodicalIF":7.2,"publicationDate":"2024-11-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142720582","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Multi-view attention graph convolutional networks for the host prediction of phages 用于噬菌体宿主预测的多视角注意力图卷积网络

IF 7.2 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Knowledge-Based Systems

Pub Date : 2024-11-22 DOI: 10.1016/j.knosys.2024.112755

Lijia Ma , Peng Gao , Wenxiang Zhou , Qiuzhen Lin , Yuan Bai , Min Fang , Zhihua Du , Jianqiang Li

Phages play pivotal roles in various biological processes, and the study of host prediction of phages (HPP) has received significant attention in recent years. HPP tries to find the specific bacteria that can be infected by certain phages, which is fundamental for the applications of targeted phage therapies and interventions. However, the existing HPP methods are mainly based on traditional wet-lab experiments which are laborious and time-consuming. Although certain computational methods have emerged to solve those issues, they perform poorly in genomes and contigs of phages as they neglect the similarity between phages in sequences and protein clusters. In this article, we propose a simple but accurate multi-view attention graph convolutional network (called PGCN) for solving the HPP problem. PGCN first constructs two phage similarity networks as a multi-view graph, which captures the similarity between phages in sequences and protein clusters. Then, PGCN uses a graph convolutional network to capture features of phages from the multi-view graph. Finally, PGCN proposes an adaptive attention mechanism to obtain the integrated features of phages from the multi-view features. Experimental results show the superiority of PGCN over the state-of-the-art methods in host prediction. The results also show the excellent performance of PGCN on host prediction in the metagenomes.

噬菌体在各种生物过程中发挥着举足轻重的作用，近年来，噬菌体宿主预测（HPP）研究受到了广泛关注。HPP 试图找到能被某些噬菌体感染的特定细菌，这对于噬菌体靶向疗法和干预措施的应用至关重要。然而，现有的 HPP 方法主要基于传统的湿实验室实验，费时费力。虽然已经出现了一些计算方法来解决这些问题，但由于它们忽视了噬菌体在序列和蛋白质簇上的相似性，因此在噬菌体的基因组和contigs上表现不佳。在本文中，我们提出了一种简单但精确的多视图注意力图卷积网络（称为 PGCN）来解决 HPP 问题。PGCN 首先将两个噬菌体相似性网络构建为多视图图，捕捉噬菌体在序列和蛋白质簇上的相似性。然后，PGCN 使用图卷积网络从多视图中捕捉噬菌体的特征。最后，PGCN 提出了一种自适应注意机制，从多视图特征中获取噬菌体的综合特征。实验结果表明，PGCN 在宿主预测方面优于最先进的方法。实验结果还显示了 PGCN 在元基因组宿主预测方面的卓越性能。

{"title":"Multi-view attention graph convolutional networks for the host prediction of phages","authors":"Lijia Ma , Peng Gao , Wenxiang Zhou , Qiuzhen Lin , Yuan Bai , Min Fang , Zhihua Du , Jianqiang Li","doi":"10.1016/j.knosys.2024.112755","DOIUrl":"10.1016/j.knosys.2024.112755","url":null,"abstract":"<div><div>Phages play pivotal roles in various biological processes, and the study of host prediction of phages (HPP) has received significant attention in recent years. HPP tries to find the specific bacteria that can be infected by certain phages, which is fundamental for the applications of targeted phage therapies and interventions. However, the existing HPP methods are mainly based on traditional wet-lab experiments which are laborious and time-consuming. Although certain computational methods have emerged to solve those issues, they perform poorly in genomes and contigs of phages as they neglect the similarity between phages in sequences and protein clusters. In this article, we propose a simple but accurate multi-view attention graph convolutional network (called PGCN) for solving the HPP problem. PGCN first constructs two phage similarity networks as a multi-view graph, which captures the similarity between phages in sequences and protein clusters. Then, PGCN uses a graph convolutional network to capture features of phages from the multi-view graph. Finally, PGCN proposes an adaptive attention mechanism to obtain the integrated features of phages from the multi-view features. Experimental results show the superiority of PGCN over the state-of-the-art methods in host prediction. The results also show the excellent performance of PGCN on host prediction in the metagenomes.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"308 ","pages":"Article 112755"},"PeriodicalIF":7.2,"publicationDate":"2024-11-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142720520","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Lightweight video object segmentation: Integrating online knowledge distillation for fast segmentation 轻量级视频对象分割：整合在线知识提炼，实现快速分割

IF 7.2 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Knowledge-Based Systems

Pub Date : 2024-11-22 DOI: 10.1016/j.knosys.2024.112759

Zhiqiang Hou , Chenxu Wang , Sugang Ma , Jiale Dong , Yunchen Wang , Wangsheng Yu , Xiaobao Yang

The typical shortcoming of STM (Space-Time Memory Network) mode video object segmentation algorithms is their high segmentation performance coupled with slow processing speeds, which poses challenges in meeting real-world application demands. In this work, we propose using an online knowledge distillation method to develop a lightweight video segmentation algorithm based on the STM mode, achieving fast segmentation while maintaining performance. Specifically, we utilize a novel adaptive learning rate to tackle the issue of inverse learning during distillation. Subsequently, we introduce a Smooth Block mechanism to reduce the impact of structural disparities between the teacher and student models on distillation outcomes. Moreover, to reduce the fitting difficulty of the student model on single-frame features, we design the Space-Time Feature Fusion (STFF) module to provide appearance and position priors for the feature fitting process of each frame. Finally, we employ a simple Discriminator module for adversarial training with the student model, to encourage the student model to learn the feature distribution of the teacher model. Extensive experiments show that our algorithm attains performance comparable to the current state-of-the-art on both DAVIS and YouTube datasets, despite running up to

\times

4 faster, with

\times

20 fewer parameters and

\times

30 fewer GFLOPS.

STM（时空记忆网络）模式视频对象分割算法的典型缺点是分割性能高，但处理速度慢，这给满足实际应用需求带来了挑战。在这项工作中，我们提出利用在线知识提炼方法开发一种基于 STM 模式的轻量级视频分割算法，在保证性能的同时实现快速分割。具体来说，我们利用一种新颖的自适应学习率来解决蒸馏过程中的反向学习问题。随后，我们引入了平滑块机制，以减少教师模型和学生模型之间的结构差异对蒸馏结果的影响。此外，为了降低学生模型在单帧特征上的拟合难度，我们设计了时空特征融合（STFF）模块，为每帧的特征拟合过程提供外观和位置先验。最后，我们采用一个简单的判别模块对学生模型进行对抗训练，鼓励学生模型学习教师模型的特征分布。大量实验表明，我们的算法在 DAVIS 和 YouTube 数据集上的性能与当前最先进的算法不相上下，尽管运行速度快了 4 倍，参数减少了 20 倍，GFLOPS 减少了 30 倍。

{"title":"Lightweight video object segmentation: Integrating online knowledge distillation for fast segmentation","authors":"Zhiqiang Hou , Chenxu Wang , Sugang Ma , Jiale Dong , Yunchen Wang , Wangsheng Yu , Xiaobao Yang","doi":"10.1016/j.knosys.2024.112759","DOIUrl":"10.1016/j.knosys.2024.112759","url":null,"abstract":"<div><div>The typical shortcoming of STM (Space-Time Memory Network) mode video object segmentation algorithms is their high segmentation performance coupled with slow processing speeds, which poses challenges in meeting real-world application demands. In this work, we propose using an online knowledge distillation method to develop a lightweight video segmentation algorithm based on the STM mode, achieving fast segmentation while maintaining performance. Specifically, we utilize a novel adaptive learning rate to tackle the issue of inverse learning during distillation. Subsequently, we introduce a Smooth Block mechanism to reduce the impact of structural disparities between the teacher and student models on distillation outcomes. Moreover, to reduce the fitting difficulty of the student model on single-frame features, we design the Space-Time Feature Fusion (STFF) module to provide appearance and position priors for the feature fitting process of each frame. Finally, we employ a simple Discriminator module for adversarial training with the student model, to encourage the student model to learn the feature distribution of the teacher model. Extensive experiments show that our algorithm attains performance comparable to the current state-of-the-art on both DAVIS and YouTube datasets, despite running up to <span><math><mo>×</mo></math></span>4 faster, with <span><math><mo>×</mo></math></span>20 fewer parameters and <span><math><mo>×</mo></math></span>30 fewer GFLOPS.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"308 ","pages":"Article 112759"},"PeriodicalIF":7.2,"publicationDate":"2024-11-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142704948","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

UrduHope: Analysis of hope and hopelessness in Urdu texts 乌尔都语希望：分析乌尔都语文本中的希望与绝望

IF 7.2 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Knowledge-Based Systems

Pub Date : 2024-11-22 DOI: 10.1016/j.knosys.2024.112746

Fazlourrahman Balouchzahi , Sabur Butt , Maaz Amjad , Grigori Sidorov , Alexander Gelbukh

Hope is a crucial aspect of human psychology that has received considerable attention due to its role in facing challenges in human life. However, current research predominantly focuses on hope as positive anticipation, overlooking its counterpart, hopelessness. This paper addresses this gap by presenting an expanded framework for analyzing hope speech in social media, incorporating hope and hopelessness. Drawing on insights from psychology and Natural Language Processing (NLP), we argue that a comprehensive understanding of human emotions necessitates considering both constructs. We introduce the concept of hopelessness as a distinct category in hope speech analysis and develop a novel dataset for Urdu, an underrepresented language in NLP research. We proposed a semi-supervised annotation procedure by utilizing Large Language Models (LLMs) along with human annotators to annotate the dataset and explored various learning approaches for hope speech detection, including traditional machine learning models, neural networks, and state-of-the-art transformers. The findings demonstrate the effectiveness of different learning approaches in capturing the nuances of hope speech in Urdu social media discourse. The hope speech detection task was modeled in two subtasks: a binary classification of Urdu tweets to Hope and Not Hope classes and then a multiclass classification of Urdu tweets into Generalized, Realistic, and Unrealistic Hopes, along with Hopelessness, and Not Hope (Neutral) categories. The best results for binary classification were obtained with Logistic Regression (LR) with an averaged macro F1 score of 0.7593, and for the multiclass classification experiments, transformers outperformed other experiments with an averaged macro F1 score of 0.4801.

希望是人类心理的一个重要方面，由于其在面对人类生活中的挑战时所发挥的作用而受到广泛关注。然而，目前的研究主要关注作为积极预期的希望，而忽略了与之相对应的绝望。本文针对这一空白，提出了一个用于分析社交媒体中希望言论的扩展框架，其中包含了希望和绝望。借鉴心理学和自然语言处理（NLP）的见解，我们认为，要全面了解人类情绪，就必须同时考虑这两种情绪。我们引入了 "无望 "这一概念，将其作为希望语音分析中的一个独特类别，并为乌尔都语（一种在 NLP 研究中代表性不足的语言）开发了一个新颖的数据集。我们提出了一种半监督注释程序，利用大型语言模型（LLMs）和人工注释员对数据集进行注释，并探索了用于希望语音检测的各种学习方法，包括传统的机器学习模型、神经网络和最先进的转换器。研究结果表明，不同的学习方法能有效捕捉乌尔都语社交媒体话语中希望语音的细微差别。希望语音检测任务分为两个子任务：将乌尔都语推文分为希望类和非希望类的二元分类，以及将乌尔都语推文分为一般希望类、现实希望类和非现实希望类以及无望类和非希望类（中性）的多类分类。在二元分类实验中，逻辑回归（LR）获得了最佳结果，平均宏 F1 得分为 0.7593；在多类分类实验中，转换器的表现优于其他实验，平均宏 F1 得分为 0.4801。

{"title":"UrduHope: Analysis of hope and hopelessness in Urdu texts","authors":"Fazlourrahman Balouchzahi , Sabur Butt , Maaz Amjad , Grigori Sidorov , Alexander Gelbukh","doi":"10.1016/j.knosys.2024.112746","DOIUrl":"10.1016/j.knosys.2024.112746","url":null,"abstract":"<div><div>Hope is a crucial aspect of human psychology that has received considerable attention due to its role in facing challenges in human life. However, current research predominantly focuses on hope as positive anticipation, overlooking its counterpart, hopelessness. This paper addresses this gap by presenting an expanded framework for analyzing hope speech in social media, incorporating hope and hopelessness. Drawing on insights from psychology and Natural Language Processing (NLP), we argue that a comprehensive understanding of human emotions necessitates considering both constructs. We introduce the concept of hopelessness as a distinct category in hope speech analysis and develop a novel dataset for Urdu, an underrepresented language in NLP research. We proposed a semi-supervised annotation procedure by utilizing Large Language Models (LLMs) along with human annotators to annotate the dataset and explored various learning approaches for hope speech detection, including traditional machine learning models, neural networks, and state-of-the-art transformers. The findings demonstrate the effectiveness of different learning approaches in capturing the nuances of hope speech in Urdu social media discourse. The hope speech detection task was modeled in two subtasks: a binary classification of Urdu tweets to Hope and Not Hope classes and then a multiclass classification of Urdu tweets into Generalized, Realistic, and Unrealistic Hopes, along with Hopelessness, and Not Hope (Neutral) categories. The best results for binary classification were obtained with Logistic Regression (LR) with an averaged macro F1 score of 0.7593, and for the multiclass classification experiments, transformers outperformed other experiments with an averaged macro F1 score of 0.4801.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"308 ","pages":"Article 112746"},"PeriodicalIF":7.2,"publicationDate":"2024-11-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142705076","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Multi-domain dialogue state tracking via dual dynamic graph with hierarchical slot selector 通过带分层插槽选择器的双动态图进行多域对话状态跟踪

IF 7.2 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Knowledge-Based Systems

Pub Date : 2024-11-22 DOI: 10.1016/j.knosys.2024.112754

Yeseul Gong, Heeseon Kim, Seokju Hwang, Donghyun Kim, Kyong-Ho Lee

Dialogue state tracking aims to maintain user intent as a consistent state across multi-domains to accomplish natural dialogue systems. However, previous researches often fall short in capturing the difference of multiple slot types and fail to adequately consider the selection of discerning information. The increase in unnecessary information correlates with a decrease in predictive performance. Therefore, the careful selection of high-quality information is imperative. Moreover, considering that the types of essential and available information vary for each slot, the process of selecting appropriate information may also differ. To address these issues, we propose HS2DG-DST, a Hierarchical Slot Selector and Dual Dynamic Graph-based DST. Our model is designed to provide maximum information for optimal value prediction by clearly exploiting the need for differentiated information for each slot. First, we hierarchically classify slot types based on the multiple properties. Then, two dynamic graphs provide highly relevant information to each slot. Experimental results on MultiWOZ datasets demonstrate that our model outperforms state-of-the-art models.

对话状态跟踪的目的是在多领域中保持用户意图的一致状态，以实现自然对话系统。然而，以往的研究往往无法捕捉到多种槽类型的差异，也没有充分考虑辨别信息的选择。不必要信息的增加会导致预测性能的下降。因此，谨慎选择高质量的信息势在必行。此外，考虑到每个时隙的基本信息和可用信息的类型各不相同，选择适当信息的过程也可能不同。为了解决这些问题，我们提出了 HS2DG-DST，一种基于分层插槽选择器和双动态图的 DST。我们的模型旨在通过明确利用每个时隙对差异化信息的需求，为最优值预测提供最大信息量。首先，我们根据多种属性对插槽类型进行分层分类。然后，两个动态图为每个插槽提供高度相关的信息。在 MultiWOZ 数据集上的实验结果表明，我们的模型优于最先进的模型。

{"title":"Multi-domain dialogue state tracking via dual dynamic graph with hierarchical slot selector","authors":"Yeseul Gong, Heeseon Kim, Seokju Hwang, Donghyun Kim, Kyong-Ho Lee","doi":"10.1016/j.knosys.2024.112754","DOIUrl":"10.1016/j.knosys.2024.112754","url":null,"abstract":"<div><div>Dialogue state tracking aims to maintain user intent as a consistent state across multi-domains to accomplish natural dialogue systems. However, previous researches often fall short in capturing the difference of multiple slot types and fail to adequately consider the selection of discerning information. The increase in unnecessary information correlates with a decrease in predictive performance. Therefore, the careful selection of high-quality information is imperative. Moreover, considering that the types of essential and available information vary for each slot, the process of selecting appropriate information may also differ. To address these issues, we propose HS2DG-DST, a Hierarchical Slot Selector and Dual Dynamic Graph-based DST. Our model is designed to provide maximum information for optimal value prediction by clearly exploiting the need for differentiated information for each slot. First, we hierarchically classify slot types based on the multiple properties. Then, two dynamic graphs provide highly relevant information to each slot. Experimental results on MultiWOZ datasets demonstrate that our model outperforms state-of-the-art models.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"308 ","pages":"Article 112754"},"PeriodicalIF":7.2,"publicationDate":"2024-11-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142720583","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

CaMo: Capturing the modularity by end-to-end models for Symbolic Regression

IF 7.2 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Knowledge-Based Systems

Pub Date : 2024-11-22 DOI: 10.1016/j.knosys.2024.112747

Jingyi Liu , Min Wu , Lina Yu , Weijun Li , Wenqiang Li , Yanjie Li , Meilan Hao , Yusong Deng , Shu Wei

Modularity is a ubiquitous principle that permeates various aspects of nature, society, and human endeavors, from biological systems to organizational structures and beyond. In the context of Symbolic Regression, which aims to find the explicit expressions from observed data, modularity could be viewed as a type of knowledge to capture the salient substructure to achieve higher fitting results. Symbolic Regression is essentially a composition optimization problem thus remaining valuable sub-structures can provide efficiency to the subsequent search. In this paper, we propose to acquire modularity in a search process and use the term module indicating the useful sub-structure. Specifically, the end-to-end model is chosen to incorporate the module into the search procedure for its scalability and generalization ability. Modules are considered high-order knowledge and act as fundamental operators, expanding the search library of Symbolic Regression. The proposed algorithm enables self-learning or self-evolution of modules as part of the learning component. Additionally, a module extraction strategy generates modules hierarchically from the expression tree, along with a module update mechanism designed to eliminate unnecessary modules while incorporating new useful ones effectively. Experiments were conducted to evaluate the effectiveness of each component.

{"title":"CaMo: Capturing the modularity by end-to-end models for Symbolic Regression","authors":"Jingyi Liu , Min Wu , Lina Yu , Weijun Li , Wenqiang Li , Yanjie Li , Meilan Hao , Yusong Deng , Shu Wei","doi":"10.1016/j.knosys.2024.112747","DOIUrl":"10.1016/j.knosys.2024.112747","url":null,"abstract":"<div><div>Modularity is a ubiquitous principle that permeates various aspects of nature, society, and human endeavors, from biological systems to organizational structures and beyond. In the context of Symbolic Regression, which aims to find the explicit expressions from observed data, modularity could be viewed as a type of knowledge to capture the salient substructure to achieve higher fitting results. Symbolic Regression is essentially a composition optimization problem thus remaining valuable sub-structures can provide efficiency to the subsequent search. In this paper, we propose to acquire modularity in a search process and use the term <em>module</em> indicating the useful sub-structure. Specifically, the end-to-end model is chosen to incorporate the module into the search procedure for its scalability and generalization ability. Modules are considered high-order knowledge and act as fundamental operators, expanding the search library of Symbolic Regression. The proposed algorithm enables self-learning or self-evolution of modules as part of the learning component. Additionally, a module extraction strategy generates modules hierarchically from the expression tree, along with a module update mechanism designed to eliminate unnecessary modules while incorporating new useful ones effectively. Experiments were conducted to evaluate the effectiveness of each component.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"309 ","pages":"Article 112747"},"PeriodicalIF":7.2,"publicationDate":"2024-11-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142748610","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A Google Trend enhanced deep learning model for the prediction of renewable energy asset price 用于预测可再生能源资产价格的谷歌趋势增强型深度学习模型

IF 7.2 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Knowledge-Based Systems

Pub Date : 2024-11-21 DOI: 10.1016/j.knosys.2024.112733

Lalatendu Mishra , Balaji Dinesh , P.M. Kavyassree , Nachiketa Mishra

This study investigates the predictive efficiency of various forecasting models for renewable energy asset prices, using oil price and investor sentiment. For renewable energy assets, renewable energy exchange-traded funds (ETFs) are considered in this study. We construct two sentiment indices using the first principal component: a fund-level investor sentiment index based on traditional indices (the Relative Strength Index and the Psychological Line Index) and the Google Trend Index derived from search trend data with keywords related to respective renewable energy ETFs. In this study, we propose a prediction model along with a deep learning framework, integrating both sentiment indices. We predict ETF log returns and conditional volatility using machine learning and deep learning models. To enhance predictive accuracy, we modify both the traditional sentiment and Google Trend indices. The results assert that models incorporating both the modified fund-level investor sentiment and Google Trends indices outperform unmodified indices. This study underscores the effectiveness of integrating multi-source sentiment for improved predictive performance, with a significant contribution by the Google Trend Index. Our model, particularly the CNN-LSTM, outperforms the CNN and BiLSTM models, as validated through Modified Diebold-Mariano tests. In addition to this benchmark, we perform additional benchmarking with forecasting techniques used in the latest ETF study and verify the robustness of our model. The findings of this study will be useful for different stakeholders of the renewable energy sector.

本研究利用石油价格和投资者情绪调查了各种可再生能源资产价格预测模型的预测效率。对于可再生能源资产，本研究考虑了可再生能源交易所交易基金（ETF）。我们利用第一个主成分构建了两个情绪指数：一个是基于传统指数（相对强弱指数和心理线指数）的基金级投资者情绪指数，另一个是根据与各可再生能源 ETF 相关关键词的搜索趋势数据得出的谷歌趋势指数。在本研究中，我们提出了一个预测模型和一个深度学习框架，将这两个情绪指数整合在一起。我们使用机器学习和深度学习模型预测 ETF 的对数收益率和条件波动率。为了提高预测准确性，我们修改了传统的情感指数和谷歌趋势指数。结果表明，包含修改后的基金级投资者情绪指数和谷歌趋势指数的模型优于未修改的指数。这项研究强调了整合多源情感以提高预测性能的有效性，其中谷歌趋势指数做出了重大贡献。我们的模型（尤其是 CNN-LSTM）优于 CNN 和 BiLSTM 模型，这一点已通过修改后的 Diebold-Mariano 测试得到验证。除此基准外，我们还对最新 ETF 研究中使用的预测技术进行了额外的基准测试，并验证了我们模型的稳健性。本研究的结果将对可再生能源领域的不同利益相关者有所帮助。

{"title":"A Google Trend enhanced deep learning model for the prediction of renewable energy asset price","authors":"Lalatendu Mishra , Balaji Dinesh , P.M. Kavyassree , Nachiketa Mishra","doi":"10.1016/j.knosys.2024.112733","DOIUrl":"10.1016/j.knosys.2024.112733","url":null,"abstract":"<div><div>This study investigates the predictive efficiency of various forecasting models for renewable energy asset prices, using oil price and investor sentiment. For renewable energy assets, renewable energy exchange-traded funds (ETFs) are considered in this study. We construct two sentiment indices using the first principal component: a fund-level investor sentiment index based on traditional indices (the Relative Strength Index and the Psychological Line Index) and the Google Trend Index derived from search trend data with keywords related to respective renewable energy ETFs. In this study, we propose a prediction model along with a deep learning framework, integrating both sentiment indices. We predict ETF log returns and conditional volatility using machine learning and deep learning models. To enhance predictive accuracy, we modify both the traditional sentiment and Google Trend indices. The results assert that models incorporating both the modified fund-level investor sentiment and Google Trends indices outperform unmodified indices. This study underscores the effectiveness of integrating multi-source sentiment for improved predictive performance, with a significant contribution by the Google Trend Index. Our model, particularly the CNN-LSTM, outperforms the CNN and BiLSTM models, as validated through Modified Diebold-Mariano tests. In addition to this benchmark, we perform additional benchmarking with forecasting techniques used in the latest ETF study and verify the robustness of our model. The findings of this study will be useful for different stakeholders of the renewable energy sector.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"308 ","pages":"Article 112733"},"PeriodicalIF":7.2,"publicationDate":"2024-11-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142720585","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

CATrack: Condition-aware multi-object tracking with temporally enhanced appearance features CATrack：利用时间增强型外观特征进行条件感知多目标跟踪

IF 7.2 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Knowledge-Based Systems

Pub Date : 2024-11-21 DOI: 10.1016/j.knosys.2024.112760

Yanchao Wang , Run Li , Dawei Zhang , Minglu Li , Jinli Cao , Zhonglong Zheng

Multiple Object Tracking (MOT) is a critical task in computer vision with a wide range of practical applications. However, current methods often use a uniform approach for associating all targets, overlooking the varying conditions of each target. This can lead to performance degradation, especially in crowded scenes with dense targets. To address this issue, we propose a novel Condition-Aware Tracking method (CATrack) to differentiate the appearance feature flow for targets under different conditions. Specifically, we propose three designs for data association and feature update. First, we develop an Adaptive Appearance Association Module (AAAM) that selects suitable track templates based on detection conditions, reducing association errors in long-tail cases like occlusions or motion blur. Second, we design an ambiguous track filtering Selective Update strategy (SU) that filters out potential low-quality embeddings. Thus, the noise accumulation in the maintained track feature will also be reduced. Meanwhile, we propose a confidence-based Adaptive Exponential Moving Average (AEMA) method for the feature state transition. By adaptively adjusting the weights of track and detection embeddings, our AEMA better preserves high-quality target features. By integrating the above modules, CATrack enhances the discriminative capability of appearance features and improves the robustness of appearance-based associations. Extensive experiments on the MOT17 and MOT20 benchmarks validate the effectiveness of the proposed CATrack. Notably, the state-of-the-art results on MOT20 demonstrate the superiority of our method in highly crowded scenarios.

多目标跟踪（MOT）是计算机视觉领域的一项重要任务，有着广泛的实际应用。然而，目前的方法通常使用统一的方法来关联所有目标，而忽略了每个目标的不同情况。这会导致性能下降，尤其是在目标密集的拥挤场景中。为了解决这个问题，我们提出了一种新颖的条件感知跟踪方法（CATrack），以区分不同条件下目标的外观特征流。具体来说，我们提出了三种数据关联和特征更新设计。首先，我们开发了一种自适应外观关联模块（AAAM），它能根据检测条件选择合适的跟踪模板，从而减少闭塞或运动模糊等长尾情况下的关联错误。其次，我们设计了一种模糊轨迹过滤选择性更新策略（SU），可以过滤掉潜在的低质量嵌入。因此，保持轨迹特征的噪声积累也会减少。同时，我们提出了一种基于置信度的自适应指数移动平均（AEMA）方法，用于特征状态转换。通过自适应调整轨迹和检测嵌入的权重，我们的 AEMA 能更好地保留高质量的目标特征。通过整合上述模块，CATrack 增强了外观特征的判别能力，提高了基于外观关联的鲁棒性。在 MOT17 和 MOT20 基准上进行的大量实验验证了所提出的 CATrack 的有效性。值得注意的是，在 MOT20 上的一流结果证明了我们的方法在高度拥挤的场景中的优越性。

{"title":"CATrack: Condition-aware multi-object tracking with temporally enhanced appearance features","authors":"Yanchao Wang , Run Li , Dawei Zhang , Minglu Li , Jinli Cao , Zhonglong Zheng","doi":"10.1016/j.knosys.2024.112760","DOIUrl":"10.1016/j.knosys.2024.112760","url":null,"abstract":"<div><div>Multiple Object Tracking (MOT) is a critical task in computer vision with a wide range of practical applications. However, current methods often use a uniform approach for associating all targets, overlooking the varying conditions of each target. This can lead to performance degradation, especially in crowded scenes with dense targets. To address this issue, we propose a novel Condition-Aware Tracking method (CATrack) to differentiate the appearance feature flow for targets under different conditions. Specifically, we propose three designs for data association and feature update. First, we develop an Adaptive Appearance Association Module (AAAM) that selects suitable track templates based on detection conditions, reducing association errors in long-tail cases like occlusions or motion blur. Second, we design an ambiguous track filtering Selective Update strategy (SU) that filters out potential low-quality embeddings. Thus, the noise accumulation in the maintained track feature will also be reduced. Meanwhile, we propose a confidence-based Adaptive Exponential Moving Average (AEMA) method for the feature state transition. By adaptively adjusting the weights of track and detection embeddings, our AEMA better preserves high-quality target features. By integrating the above modules, CATrack enhances the discriminative capability of appearance features and improves the robustness of appearance-based associations. Extensive experiments on the MOT17 and MOT20 benchmarks validate the effectiveness of the proposed CATrack. Notably, the state-of-the-art results on MOT20 demonstrate the superiority of our method in highly crowded scenarios.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"308 ","pages":"Article 112760"},"PeriodicalIF":7.2,"publicationDate":"2024-11-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142720589","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Online learning discriminative sparse convolution networks for robust UAV object tracking 在线学习判别稀疏卷积网络，实现稳健的无人飞行器目标跟踪

IF 7.2 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Knowledge-Based Systems

Pub Date : 2024-11-21 DOI: 10.1016/j.knosys.2024.112742

Qi Xu , Zhuoming Xu , Huabin Wang , Yun Chen , Liang Tao

Despite the remarkable empirical success for UAV object tracking, current convolutional networks usually have three unavoidable limitations: (1) The feature maps produced by convolutional layers are difficult to interpret. (2) The network needs to be trained offline on a large-scale auxiliary training set, resulting in the feature extraction ability of the trained network depending on the categories of the training set. (3) The performance of networks suffers from sensitivity to hyper-parameters (such as learning rate and weight decay) when the network needs online fine-tuning. To overcome the three limitations, this paper proposes a Discriminative Sparse Convolutional Network (DSCN) that exhibits good layer-wise interpretability and can be trained online without requiring any auxiliary training data. By imposing sparsity constraints on the convolutional kernels, DSCN furnishes the convolution layer with an explicit data meaning, thus enhancing the interpretability of the feature maps. These convolutional kernels are directly learned online from image blocks, which eliminates the offline training process on auxiliary data sets. Moreover, a simple yet effective online tuning method with few hyper-parameters is proposed to fine-tune fully connected layers online. We have successfully applied DSCN to UAV object tracking and conducted extensive experiments on six mainstream UAV datasets. The experimental results demonstrate that our method performs favorably against several state-of-the-art tracking algorithms in terms of tracking accuracy and robustness.

尽管卷积网络在无人机物体跟踪方面取得了显著的经验成功，但目前的卷积网络通常存在三个不可避免的局限性：（1）卷积层产生的特征图难以解释。(2）网络需要在大规模辅助训练集上进行离线训练，导致训练网络的特征提取能力取决于训练集的类别。(3) 当网络需要在线微调时，网络性能会受到超参数（如学习率和权重衰减）的影响。为了克服上述三个局限性，本文提出了一种判别稀疏卷积网络（DSCN），它具有良好的层向可解释性，并且无需任何辅助训练数据即可进行在线训练。通过对卷积核施加稀疏性约束，DSCN 为卷积层提供了明确的数据含义，从而提高了特征图的可解释性。这些卷积核直接从图像块中在线学习，从而省去了在辅助数据集上的离线训练过程。此外，我们还提出了一种简单而有效的在线调整方法，只需几个超参数就能在线微调全连接层。我们成功地将 DSCN 应用于无人机物体跟踪，并在六个主流无人机数据集上进行了广泛的实验。实验结果表明，我们的方法在跟踪精度和鲁棒性方面优于几种最先进的跟踪算法。

{"title":"Online learning discriminative sparse convolution networks for robust UAV object tracking","authors":"Qi Xu , Zhuoming Xu , Huabin Wang , Yun Chen , Liang Tao","doi":"10.1016/j.knosys.2024.112742","DOIUrl":"10.1016/j.knosys.2024.112742","url":null,"abstract":"<div><div>Despite the remarkable empirical success for UAV object tracking, current convolutional networks usually have three unavoidable limitations: (1) The feature maps produced by convolutional layers are difficult to interpret. (2) The network needs to be trained offline on a large-scale auxiliary training set, resulting in the feature extraction ability of the trained network depending on the categories of the training set. (3) The performance of networks suffers from sensitivity to hyper-parameters (such as learning rate and weight decay) when the network needs online fine-tuning. To overcome the three limitations, this paper proposes a Discriminative Sparse Convolutional Network (DSCN) that exhibits good layer-wise interpretability and can be trained online without requiring any auxiliary training data. By imposing sparsity constraints on the convolutional kernels, DSCN furnishes the convolution layer with an explicit data meaning, thus enhancing the interpretability of the feature maps. These convolutional kernels are directly learned online from image blocks, which eliminates the offline training process on auxiliary data sets. Moreover, a simple yet effective online tuning method with few hyper-parameters is proposed to fine-tune fully connected layers online. We have successfully applied DSCN to UAV object tracking and conducted extensive experiments on six mainstream UAV datasets. The experimental results demonstrate that our method performs favorably against several state-of-the-art tracking algorithms in terms of tracking accuracy and robustness.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"308 ","pages":"Article 112742"},"PeriodicalIF":7.2,"publicationDate":"2024-11-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142704949","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Enhancing learning process modeling for session-aware knowledge tracing 加强学习过程建模，实现会话感知知识追踪

IF 7.2 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Knowledge-Based Systems

Pub Date : 2024-11-21 DOI: 10.1016/j.knosys.2024.112740

Chunli Huang , Wenjun Jiang , Kenli Li , Jie Wu , Ji Zhang

Session-aware knowledge tracing tries to predict learners’ performance, by splitting learners’ sequences into sessions and modeling their learning within and between sessions. However, there still is a lack of comprehensive understanding of the learning processes and session-form learning patterns. Moreover, the knowledge state shifts between sessions at the knowledge concept level remain unexplored. To this end, we conduct in-depth data analysis to understand learners’ learning processes and session-form learning patterns. Then, we perform an empirical study validating knowledge state shifts at the knowledge concept level in real-world educational datasets. Subsequently, a method of Enhancing Learning Process Modeling for Session-aware Knowledge Tracing, ELPKT, is proposed to capture the knowledge state shifts at the knowledge concept level and track knowledge state across sessions. Specifically, the ELPKT models learners’ learning process as intra-sessions and inter-sessions from the knowledge concept level. In intra-sessions, fine-grained behaviors are used to capture learners’ short-term knowledge states accurately. In inter-sessions, learners’ knowledge retentions and decays are modeled to capture the knowledge state shift between sessions. Extensive experiments on four real-world datasets demonstrate that ELPKT outperforms the existing methods in learners’ performance prediction. Additionally, ELPKT shows its ability to capture the knowledge state shifts between sessions and provide interpretability for the predicted results.

会话感知知识追踪试图通过将学习者的序列分割成会话，并对他们在会话内和会话间的学习进行建模，来预测学习者的成绩。然而，人们对学习过程和会话形式的学习模式仍然缺乏全面的了解。此外，在知识概念层面上，会话之间的知识状态转变仍未得到探索。为此，我们进行了深入的数据分析，以了解学习者的学习过程和会话形式的学习模式。然后，我们进行了一项实证研究，在真实世界的教育数据集中验证了知识概念层面的知识状态转移。随后，我们提出了一种增强会话感知知识追踪的学习过程建模方法（ELPKT），以捕捉知识概念层面的知识状态转变，并追踪跨会话的知识状态。具体来说，ELPKT从知识概念层面将学习者的学习过程分为会话内和会话间两个阶段。在会话内，使用细粒度行为来准确捕捉学习者的短期知识状态。在会话间，对学习者的知识保留和衰减进行建模，以捕捉会话间知识状态的转变。在四个真实世界数据集上进行的广泛实验证明，ELPKT 在学习者成绩预测方面优于现有方法。此外，ELPKT 还展示了其捕捉会话间知识状态转变的能力，并为预测结果提供了可解释性。

{"title":"Enhancing learning process modeling for session-aware knowledge tracing","authors":"Chunli Huang , Wenjun Jiang , Kenli Li , Jie Wu , Ji Zhang","doi":"10.1016/j.knosys.2024.112740","DOIUrl":"10.1016/j.knosys.2024.112740","url":null,"abstract":"<div><div>Session-aware knowledge tracing tries to predict learners’ performance, by splitting learners’ sequences into sessions and modeling their learning within and between sessions. However, there still is a lack of comprehensive understanding of the learning processes and session-form learning patterns. Moreover, the knowledge state shifts between sessions at the knowledge concept level remain unexplored. To this end, we conduct in-depth data analysis to understand learners’ learning processes and session-form learning patterns. Then, we perform an empirical study validating knowledge state shifts at the knowledge concept level in real-world educational datasets. Subsequently, a method of Enhancing Learning Process Modeling for Session-aware Knowledge Tracing, ELPKT, is proposed to capture the knowledge state shifts at the knowledge concept level and track knowledge state across sessions. Specifically, the ELPKT models learners’ learning process as intra-sessions and inter-sessions from the knowledge concept level. In intra-sessions, fine-grained behaviors are used to capture learners’ short-term knowledge states accurately. In inter-sessions, learners’ knowledge retentions and decays are modeled to capture the knowledge state shift between sessions. Extensive experiments on four real-world datasets demonstrate that ELPKT outperforms the existing methods in learners’ performance prediction. Additionally, ELPKT shows its ability to capture the knowledge state shifts between sessions and provide interpretability for the predicted results.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"309 ","pages":"Article 112740"},"PeriodicalIF":7.2,"publicationDate":"2024-11-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142720444","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0