Journal of Shaanxi Normal University(Natural Science Edition)

  • Double triadic thinking and the 3×3 methods

    SUO Langwangqing;YANG Hailong;YANG Han;YAO Yiyu;School of Mathematics and Statistics, Shaanxi Normal University;Department of Computer Science, University of Regina;College of Mathematics, Southwest Jiaotong University;

    When dealing with complex issues, developing comprehensive strategies, or making critical decisions, it is necessary to think and analyze from multiple perspectives, levels, or dimensions. Three-way decision is triadic thinking, triadic method, and triadic computing. Following the principles of three-way decision, the concept of double triadic thinking and the associated 3×3 methods and structures are introduced.Double triadic thinking is based on a combination of two triadic structures, which allows us to think, analyze and solve problems from nine different perspectives or dimensions.Two particular 3×3 methods are proposed by combining trilevel hierarchical thinking and triangular thinking. One is the application of triangular methods at each of the three levels of a hierarchy, which is called a(3-level)×(3-angle) method. The other is the application of trilevel methods at each of the three vertices of a triangle, which is called a(3-angle)×(3-level) method. As a case study, 3×3 methods are applied to explainable artificial intelligence. By means of the concept of Symbols-Meaning-Value(SMV) space, we consider specific semantics of the nine elements of a 3×3 method. The SMV space based 3×3 method can analyze and interpret the data, assumptions, principles, and outcomes of an intelligent system at multiple levels. It provides a construction process and structure of explanation for intelligent systems, making an explanation easier to communicate, understand, and accept.

    2024 03 v.52;No.213 [Abstract][OnlineView][Download 1276K]

  • A review of three-way decisions based on machine learning

    LIU Dun;GAO Luyue;School of Economics and Management, Southwest Jiaotong University;Service Science and Innovation Key Laboratory of Sichuan Province;

    Three-way decision(3WD) is an important research direction in the field of granular computing, which conforms to human thinking and can effectively deal with uncertainty problems in practical decision-making processes. Three-way decision can effectively reduce the cost of decision-making, enhance the control of uncertain decision-making processes and improve the interpretability of model by introducing delay strategy. Therefore, a three-way decision based on machine learning is worth researching deeply. Firstly, the basic model of three-way decision is introduced.Secondly, the research status of three-way decision based on machine learning at domestic and foreign search results is analyzed by using CiteSpace and VOSviewer. Thirdly, from the perspectives of research questions, model methods, and application backgrounds, the integration of three-way decisions with clustering models, classification models, recommendation systems, and deep learning models are focused on.Finally, the future research directions of three-way decisions based on machine learning are given.

    2024 03 v.52;No.213 [Abstract][OnlineView][Download 2351K]

  • A three-way risk rating method integrating expert domain knowledge and K-means clustering

    DUAN Weiyi;LIANG Decui;School of Management and Economics, University of Electronic Science and Technology of China;

    In practical domains such as finance and healthcare, decision-making problems necessitate through the consideration of risks, where precise prediction and accurate risk classification hold crucial significance. Nevertheless, traditional group decision-making studies prioritize the consistency and consensus of expert evaluations while allocating lesser attention to acquiring objective evaluations and the decision quality. Consequently, a data-driven approach is introduced to assist experts in discovering evaluation through data and clustering results, optimizing group opinions within the three-way decision framework so as to improve and calculate the discriminative point of logistic regression for the results of risk rating classification. The risk rating is determined based on four publicly available datasets of credit risk and disease diagnosis from UCI and Kaggle. Empirical results from data experiments indicate that our proposed three-way classification method focuses more on risk avoidance compared to classical machine learning methods, and achieves stable and superior performance across all datasets. This implies that utilizing objective information from data to assist expert evaluations in risk assessment can help to solve decision problems within different domains.

    2024 03 v.52;No.213 [Abstract][OnlineView][Download 1124K]

  • Three-way decision models based on interval-valued hesitant fuzzy multi-granularity rough set

    XU Weihua;DING Yi;DENG Biao;ZHANG Xiaoyan;College of Artificial Intelligence, Southwest University;

    In this paper, two three-way decision models based on interval-valued hesitant fuzzy multi-granularity rough set are proposed for the interval-valued hesitant fuzzy information system. First, the optimistic and pessimistic interval-valued hesitant fuzzy multi-granularity rough set models are determined by means of multi-granularity rough set theory. Then, the concept of interval valued hesitant fuzzy continuous cross-entropy is introduced to calculate the conditional probabilities under different cases through the technique for order preference by similarity to an ideal solution(TOPSIS). Based on this, interval-valued hesitant fuzzy decision-theoretic rough sets and relevant three-way decision rules are proposed. Finally, it is illustrated that these models adopt different attitudes and decision-making schemes for target evaluation through an example, and the effectiveness of the algorithm is verified.

    2024 03 v.52;No.213 [Abstract][OnlineView][Download 1011K]

  • Density peak clustering algorithm optimized with local standard deviation

    XIE Juanying;ZHANG Wenjie;School of Computer Science, Shaanxi Normal University;

    DPC(clustering by fast search and find of density peaks) algorithm is a density based clustering algorithm. It is one of the milestone clustering algorithms. It can find any arbitrary shapes of clusters embedded within any dimensional spaces. However, its local density definition of a point is not appropriate for simultaneously detecting the cluster centers of dense and sparse clusters, nor detecting the sparse and dense clusters subsequently. In addition, its one-step assignment strategy leads to a fatal problem, that is, once a point is assigned to an incorrect cluster, there are more subsequent points being assigned erroneously, resulting in the domino effect.To address the aforementioned problems, this paper redefines the local density of a point based on the local standard deviation, and proposes a two-step assignment strategy, resulting in the ESDTS-DPC algorithm. The ESDTS-DPC algorithm is compared with the original DPC and its variations including KNN-DPC, FKNN-DPC, DPC-CE and the classic density based clustering algorithm, such as DBSCAN. The extensive experiment results demonstrate superiority of the proposed ESDTS-DPC in detecting the clustering within a dataset.

    2024 03 v.52;No.213 [Abstract][OnlineView][Download 4202K]

  • A classification method for low-grade glioma based on gene attention and multi-omics

    CHENG Hao;HAN Xiao;REN Jianxue;YAN Aoyu;WANG Huiqing;College of Computer Science and Technology(College of Data Science),Taiyuan University of Technology;

    Existing studies on the three-class classification of molecular subtypes of low-grade glioma(LGG) rely on LGG medical imaging data. The scarcity and difficulty of obtaining data samples make it challenging for models to learn the differences between LGG molecular subtypes, reducing the model's classification performance.A three-class classification method for LGG molecular subtypes called MODDA is proposed, which utilizes a gene attention network to extract important features from LGG multi-omics data and employs an embedding network to process clinical data to obtain clinical data features. Then fuses clinical data features with important omics data features and uses a dense deep neural network for the classification of LGG molecular subtypes. Experimental results show that MODDA's classification performance surpasses existing LGG molecular subtype classification methods and also exhibits good generalization performance on external validation datasets. Moreover, an enrichment analysis of important genes identified during the chi-square testing process for gene ontology(GO) terms and biological pathways is conducted, aiding in the personalized treatment of LGG.

    2024 03 v.52;No.213 [Abstract][OnlineView][Download 2042K]

  • Prediction of protein-ligand binding affinity based on LSTM and attention mechanism

    WANG Wei;WU Shiyu;LIU Dong;LIANG Huiru;SHI Jinling;ZHOU Yun;ZHANG Hongjun;WANG Xianfang;College of Computer and Information Engineering, Henan Normal University;Key Laboratory of Artificial Intelligence and Personalized Learning in Education of Henan Province;International School of Education, Xuchang University;Hebi Instiute of Engineering and Technology, Henan Polytechnic University;College of Computer Science and Technology Engineering, Henan Institute of Technology;

    Protein-ligand binding affinity prediction is a challenging task in drug repositioning regression. Deep learning methods can effectively predict the binding affinity of protein-ligand interactions, reducing the time and cost of drug discovery. This study proposes a deep convolutional neural network model(DLLSA) based on long short-term memory module(LSTM) and attention mechanism module.The model is constructed using a convolutional network parallel pattern embedded with LSTM and spatial attention module. The LSTM module focuses on the long sequence information of protein ligand contact features, while the spatial attention module aggregates local information of contact features. PDBbind(v. 2020) dataset was used for training, and CASF-2013 and CASF-2016 datasets were used for validating. Pearson correlation coefficients of the model were improved by 0.6% and 3% compared to the PLEC model, and the experimental results were significantly better than the current correlation methods.

    2024 03 v.52;No.213 [Abstract][OnlineView][Download 1926K]

  • A novel algorithm based on the improved YOLOv7 for detecting transmission tower base

    LEI Lei;WEI Xiaolong;LIANG Jun;DONG Qian;XIAO Zhangshu;State Grid Shaanxi Electric Power Co.LTD., Electric Power Research Institute;State Grid (Xi'an) Environmental Protection Technology Center Co.LTD;State Grid Shaanxi Electric Power Co.LTD;School of Computer Science, Shaanxi Normal University;

    The pylon is one of the most important components in the entire power transmission system.It is necessary to timely inspect the tower to ensure the stability of the base for the later use. There are problems of the transmission tower images collected by UAV have complex backgrounds, the background is similar to the base of target tower, as well as small objects and incomplete tower base, this paper proposes an improved YOLOv7 algorithm for detecting the base of tower. Firstly, using the pylon images of different landforms to construct high-quality data sets. Then CBAM attention mechanism is added to the Backbone layer of the original YOLOv7 to improve the feature extraction ability of the pylon. Finally, introducing WIoU v3 instead of the original coordinate loss function CIoU to improve the veracity and stability of target detection tasks. On this dataset, a comparative experiment was conducted using the improved YOLOv7 algorithm and the current mainstream object detection algorithm. The mAP value of our algorithm is as high as 99.93% in the experimental results, it is 2.19% higher than the original YOLOv7, the FPS value is 37.125, which meets the real-time detection requirements, and the overall performance of the algorithm is good. It's feasible and effective in detection tasks of towers' base for our algorithm, which has been proven by the experiments in this paper, and laying the foundation for future research on the soil and water around the base of tower.

    2024 03 v.52;No.213 [Abstract][OnlineView][Download 2515K]

  • Research on sentence intention recognition and slot filling based on multi-task distillation

    GAO Zixiong;JIANG Shengyi;OU Yanmei;XUAN Zhenyu;School of Information Science and Technology/School of Cyber Security, Guangdong University of Foreign Studies;

    At present, pre-trained models such as BERT have achieved good results in many NLP tasks, but the pre-trained models are difficult to deploy in small configuration environments because of their large parameter scale, large computation and high requirements on hardware resources. Model compression is the key to solve this problem, and knowledge distillation is currently a better model compression method.A joint model of sentence intent recognition and slot filling based on multi-task distillation is proposed.The model applies ALBERT to task-based dialogue system, and uses the knowledge distillation strategy to migrate the ALBERT model knowledge to the BiLSTM model. Experimental results show that the sentence accuracy rate of the ALBERT based joint model in the SMP 2019 evaluation data set is 77.74%, the sentence accuracy rate of the BiLSTM model trained separately is 58.33%, and the sentence accuracy rate of the distillation model is 67.22%, which is 8.89% higher than the BiLSTM model while offering an inference speed approximately 18.9 times faster than ALBERT.

    2024 03 v.52;No.213 [Abstract][OnlineView][Download 1069K]

  • Click-through rate prediction method based on local influence and deep preference propagation

    XU Wei;LI Xiang;ZHU Quanyin;REN Ke;SUN Jizhou;School of Computer and Software Engineering, Huaiyin Institute of Technology;

    Based on knowledge graphs, recommendation methods have become one of the hot research topics in recommender systems. They utilize user historical behaviors and item features with the assistance of structured representation of data in knowledge graphs to address the sparsity and cold-start problems in recommendation systems. However, user interests are easily influenced by specific items, while knowledge graphs store data in structured forms with entities linked only through relational edges. This leads to poor performance in click-through rate prediction when solely relying on knowledge graph-based recommendation methods.A recommendation method called local influence and deep preference propagation(LIDP) is proposed, which fully exploits the advantages of entity influence in preference propagation within structured data of knowledge graphs. The LIDP model first propagates preferences layer by layer in the knowledge graph to obtain data influence weights, and then calculates local influence based on these weights. Next, it enhances user representations based on the enhanced representation of user interests through their historical behaviors. Finally, it calculates the final interaction probability by taking the inner product of user representations and item vector representations. On the MovieLens-1M dataset, compared to the optimal baseline model GNRF, LIDP improves AUC, ACC, MAE, and F_1 score by 0.16%, 0.52%, 0.87%, and 0.21% respectively. On the Book-Crossing dataset, these improvements are 0.45%, 2.14%, 1.29%, and 0.93% respectively. Experimental results demonstrate that the LIDP model effectively captures deep-level user interest preferences, exhibiting good performance and effectiveness in recommendation systems, thereby providing users with better personalized recommendation services.

    2024 03 v.52;No.213 [Abstract][OnlineView][Download 1821K]

  • Certificate image tampering detection based on contextual semantic information

    LI Pei;WANG Wei;LIU Yong;WANG Yi;Chongqing Ant Consumer Financial Co. Ltd.,Ant Group;

    During consumer financial services, there is a headache with overdue repayment. When negotiating with customers, a small number of them try to use tampered certificate images to achieve illegal benefits. These tampering focuses on content with strong contextual semantic connections such as personal information, seals, and issuing units. Based on the traditional spatial domain RGB and frequency domain DCT as discriminative features, the position of text blocks, seal blocks, and deconvolution network to realize an end-to-end fully convolutional neural network that includes semantic relations are introduced. Compared with the traditional models, it has a 3.97% higher mIoU in “Tianchi's 2022 Real Scene Tampering Image Detection Challenge” dataset. In our service scenario, the accuracy of tampering detection has been improved by 3.7%.

    2024 03 v.52;No.213 [Abstract][OnlineView][Download 1323K]