模式识别与人工智能
Home      About Journal      Editorial Board      Instructions      Ethics Statement      Contact Us                   中文
Pattern Recognition and Artificial Intelligence
22 Judgement and Disposal of Academic Misconduct Article
22 Copyright Transfer Agreement
22 Proof of Confidentiality
22 Requirements for Electronic Version
More....
22 Chinese Association of Automation
22 National ResearchCenter for Intelligent Computing System
22 Institute of Intelligent Machines,Chinese Academy of Sciences
More....
 
 
2022 Vol.35 Issue.2, Published 2022-02-25
Papers and Reports    Researches and Applications   
   
Papers and Reports
95 Modeling Method of Knowledge Relevance Based on Fuzzy Measures
ZHANG Suojuan, HUANG Song, YU Xiaohan, CHEN Enhong
The relevance between knowledge in the instructional scenarios draws much attention. The existing research usually focuses on modeling the relationship between two knowledge points. However, the complex relevance in knowledge sets is ignored, which results in the deviation of the research results. Aiming at this problem, the fuzzy measure is introduced to quantify the knowledge set, and then a modeling method of knowledge relevance based on fuzzy measures is proposed. Firstly, three different knowledge relationships are analyzed grounded on the cognitive theory, and the knowledge relevance is modeled with fuzzy measures. Then, the practicability of the modeling method is demonstrated by the practical scenario. Secondly, based on fuzzy measure modeling, the importance and interaction of knowledge are discussed from the perspective of knowledge relevance. Finally, the application of knowledge relevance in cognitive diagnosis is studied. The influence of knowledge relevance on cognitive diagnosis is demonstrated through the experiments on real-world datasets. The results show that the proposed method predicts precisely with better interpretability.
2022 Vol. 35 (2): 95-105 [Abstract] ( 758 ) [HTML 1KB] [ PDF 820KB] ( 615 )
106 AdaBelief Based Heavy-Ball Momentum Method
ZHANG Zedong, LONG Sheng, BAO Lei, TAO Qing
Adaptive moment estimation algorithms with momentum and adaptive step techniques are widely applied in deep learning. However, these algorithms cannot achieve the optimal performance in both theory and experiment. To solve the problem, an AdaBelief based heavy-ball momentum method, AdaBHB, is proposed. The AdaBelief technique of adjusting step size flexibly is introduced to improve the algorithm performance in experiments. The heavy ball momentum method with step size adjusted by exponential moving average strategy is employed to accelerate convergence. According to the convergence analysis techniques of AdaBelief and Heavy-ball momentum methods, time-varying step size and momentum coefficient are selected skillfully and the momentum term and adaptive matrix are added. It is proved that AdaBHB gains the optimal individual convergence rate for non-smooth general convex optimization problems. Finally, the correctness of the theoretical analysis of the proposed algorithm is verified by experiments on convex optimization problems and deep neural networks, and AdaBHB is validated to obtain the optimal convergence in theory with performance improved.
2022 Vol. 35 (2): 106-115 [Abstract] ( 456 ) [HTML 1KB] [ PDF 791KB] ( 566 )
116 Circular Convolutional Neural Networks Based on Triplet Attention
WANG Jingbin, LEI Jing, ZHANG Jingxuan, SUN Shounan
In the existing knowledge completion models with textual or neighbor information, the interaction between texts and neighbors is ignored. Therefore, it is difficult to capture the information with strong semantic relevance to entities. In addition, the relationship-specific information in the entities is not taken into account in the models based on convolutional neural networks, which results in poor prediction performance. In this paper, a circular convolutional neural network model based on triplet attention is proposed combining textual and neighbor information. Firstly, the words with strong semantic relevance to entities in textual descriptions are selected by semantic matching, and then they are combined with topological neighbors as entity neighbors to enhance entity representations. Next, the fusion representations of the entity and the relation representations are reshaped. Finally, the triplet attention is utilized to optimize the input of the convolution and the convolution operation can extract the features related to the relations in the entities, which improves the model performance. Experiments on several public datasets show that the performance of the proposed model is superior.
2022 Vol. 35 (2): 116-129 [Abstract] ( 500 ) [HTML 1KB] [ PDF 931KB] ( 547 )
130 Pixel?Level Segmentation Algorithm Combining Depth Map Clustering and Object Detection
FANG Baofu, ZHANG Xu,WANG Hao
Acquiring semantic information in the surrounding environment is an important task of semantic simultaneous localization and mapping(SLAM). However, the time performance of the system is affected by semantic segmentation or instance segmentation, and the accuracy of the system is reduced while adopting object detection methods. Therefore, a pixel?level segmentation algorithm combining depth map clustering and object detection is proposed in this paper. The positioning accuracy of the current semantic SLAM system is improved with the real?time performance of the system guaranteed. Firstly, the mean filtering algorithm is utilized to repair the invalid points of the depth map and thus the depth information is more reliable. Secondly, object detection is performed on RGB images and K?means clustering is employed for corresponding depth maps, and then the pixel?level object segmentation result is obtained by combining the two results. Finally, the dynamic points in the surrounding environment are eliminated by the results described above, and a complete semantic map without dynamic objects is established. Experiments of depth map restoration, pixel?level segmentation, and comparison between the estimated camera trajectory and the real camera trajectory are carried out on TUM dataset and real home scenes. The experimental results show that the proposed algorithm exhibits good real?time performance and robustness.
2022 Vol. 35 (2): 130-140 [Abstract] ( 606 ) [HTML 1KB] [ PDF 3227KB] ( 508 )
Researches and Applications
141 3D Point Cloud Classification Based on Local-Nonlocal Interactive Convolution
LU Xinyu, YANG Bing, YE Hailiang, CAO Feilong
Now 3D point cloud classificaiton is widely applied in many domains, including robot operation, automous driving and virtual reality. Extracting rich features with high discrimination is the key to 3D point cloud classification. Therefore, an algorithm of 3D point cloud classification based on local-nonlocal interactive convolution is designed to improve the feature extraction of point cloud. Firstly, a local-nonlocal interactive convolution module is constructed. After obtaining local and nonlocal similar features, interactive enhancement is employed to alleviate the redundancy problem caused by a single neighborhood representing a closed region. Consequently, the hierarchy and stability of the network are enhanced and the degradation problem of the network is alleviated. Then, the convolution neural network is constructed with the module as the basic unit. Finally, adaptive feature fusion is adopted to make full use of different levels of features to realize 3D point cloud classification. Experimental results on two benchmark datasets, ModelNet40 and ScanObjectNN, show that the proposed method generates better performance.
2022 Vol. 35 (2): 141-149 [Abstract] ( 412 ) [HTML 1KB] [ PDF 962KB] ( 434 )
150 Feature Selection Based on Adaptive Whale Optimization Algorithm and Fault-Tolerance Neighborhood Rough Sets
SUN Lin, HUANG Jinxu, XU Jiucheng, MA Yuanyuan
Traditional whale optimization algorithm(WOA) cannot handle continuous data effectively, and the tolerance of neighborhood rough sets(NRS) for noise data is poor. To address the issues, an algorithm of feature selection based on adaptive WOA and fault-tolerance NRS is presented. Firstly, a piecewise dynamic inertia weight based on iteration cycle is proposed to prevent the WOA from falling into local optimum prematurely. The shrinkage enveloping and spiral predation behaviors of WOA are improved, and an adaptive WOA is designed. Secondly, the ratio of the same decision features in the neighborhood is introduced to make up for the fault tolerance lack of NRS model for noise data, and the upper and lower approximations, approximation precision and approximation roughness, fault-tolerance dependence and approximation conditional entropy of fault-tolerance neighborhood are defined. Finally, a fitness function is constructed based on the fault-tolerance NRS, and then the adaptive WOA searches for the optimal feature subset through continuous iterations. The Fisher score is employed to reduce the dimensions of high-dimensional datasets preliminarily and the time complexity of the proposed algorithm effectively. The proposed algorithm is tested on 8 low-dimensional UCI datasets and 6 high-dimensional gene datasets. Experimental results demonstrate that the proposed algorithm selects fewer features effectively with high classification accuracy.
2022 Vol. 35 (2): 150-165 [Abstract] ( 462 ) [HTML 1KB] [ PDF 883KB] ( 644 )
166 Voting Prediction Model Based on Voter Influence Factor
ZHANG Xinyun, ZHANG Shaowu, REN Lu, YANG Liang, XU Bo, ZHANG Yijia, LIN Hongfei
Voting prediction is one of the applications of computational politics. However, the interaction between voters in the voting is ignored by most of the prediction models. To solve this problem, a voting prediction model based on voter influence factor is proposed in this paper. Firstly, the voter influence factor is proposed to describe the influence of a voter on the voting choices of other voters in the voting process. A factor graph is generated by combining the voter influence factor and the voter characteristics extracted by the pre-training model. Then, the factor graph is introduced into graph convolution neural network to learn the interaction of voters and to simulate the real voting game to a certain extent. Considering the relevance of context information in the text of the bill, bi-directional long short-term memory is utilized to obtain the feature vector of bill. In view of similar writing and repetition of words caused by standardization of the bill text, the key words of the bill are obtained by TextRank with term-frequency-inverse document frequency factor. Finally, experiments on the foreign congress website dataset show that the performance of the proposed model is superior. The ablation experiments verify that each module improves the performance of the model to a certain extent.
2022 Vol. 35 (2): 166-174 [Abstract] ( 406 ) [HTML 1KB] [ PDF 936KB] ( 436 )
175 Label-Guided Dual-Attention Deep Neural Network Model
PENG Zhanwang, ZHU Xiaofei, GUO Jiafeng
Since the text information of labels is not included in some datasets, the semantic relationship between text words and labels cannot be explicitly calculated in the existing explicit interactive classification models. To solve this problem, a label-guided dual-attention deep neural network model is proposed in this paper. Firstly, an automatic category label description generation method based on inverse label frequency is proposed. According to the label description generation method, a specific label description for each label is generated. The generated specific label description is applied to explicitly calculate the semantic relationship between text words and labels. On the basis of the above, review text representation with contextual information is learned by a text encoder. A label-guided dual-attention network is proposed to learn the text representation based on self-attention and the text representation based on label attention, respectively. Then, an adaptive gating mechanism is employed to fuse two mentioned text representations and the final text representation is thus obtained. Finally, a two-layer feedforward neural network is utilized as a classifier for sentiment classification. Experiments on three publicly available real-world datasets show that the proposed model produces better classification performance.
2022 Vol. 35 (2): 175-184 [Abstract] ( 491 ) [HTML 1KB] [ PDF 669KB] ( 359 )
185 Multi-label Classification of Legal Text with Fusion of Label Relations
SONG Zeyu, LI Yang, LI Deyu, WANG Suge

With the rapid development of big data technology, multi-label text classification spawns many applications in the judicial field. There are multiple element labels in legal texts, and the labels are interdependent or correlated. Accurate identification of these labels requires the support of multi-label classification method. In this paper, a multi-label classification method of legal texts with fusion of label relations(MLC-FLR) is proposed. A graph convolution network model is utilized to capture the dependency relationship between labels by constructing the co-occurrence matrix of labels. The label attention mechanism is employed to calculate the degrees of correlation between a legal text and each label word, and the legal text semantic representation of a specific label can be obtained. Finally, the comprehensive representation of a text for multi-label classification is carried out by combining the dependency relationship and the legal text semantic representation of a specific label. Experimental results on the legal text datasets show that MLC-FLR achieves better classification accuracy and stability.
2022 Vol. 35 (2): 185-192 [Abstract] ( 552 ) [HTML 1KB] [ PDF 727KB] ( 535 )
模式识别与人工智能
 

Supervised by
China Association for Science and Technology
Sponsored by
Chinese Association of Automation
NationalResearchCenter for Intelligent Computing System
Institute of Intelligent Machines, Chinese Academy of Sciences
Published by
Science Press
 
Copyright © 2010 Editorial Office of Pattern Recognition and Artificial Intelligence
Address: No.350 Shushanhu Road, Hefei, Anhui Province, P.R. China Tel: 0551-65591176 Fax:0551-65591176 Email: bjb@iim.ac.cn
Supported by Beijing Magtech 　Email:support@magtech.com.cn