【作者】 马超；

【导师】 欧阳继红；

【作者基本信息】 吉林大学 ， 计算机应用技术， 2014， 博士

【摘要】 分类问题是模式识别、数据挖掘和机器学习领域中最重要的研究内容之一，在现实世界有着广泛的应用。人工神经网络由于自身良好的特性，能通过自主学习从数据中获取知识信息，为分类问题提供了有效的解决途径。但是传统基于梯度下降的网络模型、支持向量网络存在训练时间过长，收敛速度过慢，易过拟合等不足，对于实际复杂分类问题，直接利用传统的分类学习方法难以得到理想的结果，如何设计出高效且泛化能力强的分类模型，是目前仍未很好解决的问题。本论文的研究工作主要围绕基于极限学习机方法设计具有良好泛化能力和分类性能的算法和模型，并用于解决医疗诊断和风险评估领域中的分类决策问题。分别提出了结合人工蜂群算法和极限学习机的优化分类算法、基于自适应人工蜂群算法智能优化多核极限学习机的分类算法、集成局部Fisher判别分析特征提取机制和进化核极限学习机方法的甲状腺疾病诊断模型、集成优化减聚类加权算法和核极限学习机的帕金森疾病诊断混合模型、基于改进引力搜索算法框架优化核极限学习机的分类方法以及信用风险评估模型。在知名UCI公共分类数据集上的实验结果表明，本论文中所提出的这些分类方法均获得良好的分类性能，其效果在很大程度上显著优于已有方法和相似算法，达到了预期的效果和目的。更多还原

【Abstract】 The Classification problem is one of the most important research topics in the areas ofpattern recognition, data mining and machine learning. The neural networks methods as animportant branch of topic, especially, the proposed Back Propagation (BP) algorithmpromotes the rapid development of neural networks in the fields of theoretical research andtechnology application. Neural networks methods can effectively find the information andpatterns from the data by independent learning, which provide an effective way to solve theseclassification problems. However, traditional neural networks such as BP algorithm andSupport Vector Networks both cost much training time, converge with a slow speed and trapin local optimum easily, how to design a classification method with validity and goodgeneralization ability, and to provide support for the scientific research and technicalapplication, that is a difficult problem need to be well solved.This paper focuses on a novel classification method called Extreme Learning Machine(ELM), which has been proposed and studied recently in the designing and construction ofclassification methods with effectiveness and good generaliza tion ability. We have proposedArtificial Bee Colony (ABC) algorithm based ELM classification method. The self-adaptiveABC method based multi-kernel ELM classification method. The local fisher discriminateanalysis method based improved kernelized ELM system for thyroid disease diagnosis model.The adaptive subtract clustering feature weighting (ASCFW) method based kernelized ELMmodel for Parkinson’s disease diagnosis model. The improved gravitational search algorithmbased kernelized ELM method for classification problems.The contributions and innovations of this paper are briefly given as follows:(1) We make a brief discussion on theoretical methods of neural networks, and make ananalysis and discussion on the development trend and defect of these methods. Moreover,we analyze and discuss the classification principle, the research status and the defect ofELM methods in detail. This part has laid the foundation for the next research.(2) ABC-based ELM method for classificationThe performance of ELM algorithm depends on the input weights and biases values of neural networks, so we use ABC algorithm to optimize the ELM parameters includinginput weights and biases. The superior of ABC algorithm with the global search abilitycan conduct ELM model to train and test, and considering the minimized output weightnorm in maximizing the classification accuracy at the same time. Experiment results showthat the proposed method can effectively improve the generalization performance of ELM,and the structure of network is more compact. The performance of kernel ELM is alsoaffected by different types of kernel functions, we propose the self-adaptive ABCalgorithm to optimize the multi-kernel ELM model. In this method, multiple kernel designcan be more flexible to reflect different data structures, SABC algorithm with four searchstrategies is used to adaptively optimize the related parameters of multi-kernel ELM forclassification. The experimental results show that the proposed method can achieve bettergeneralization performance than existing and similar methods in sixteen classificationdatasets, the construction of kernel function is flexible and reasonable, and the obtainedresults are more stable.(3) Feature extraction based improved kernel ELM model for Thyroid diagnosisThe related parameters of kernel ELM are important factors for improving theperformance of ELM methods. We propose a novel system that combines LFDA andimproved kernel ELM algorithm for Thyroid disease diagnosis. In this model, firstly,LFDA algorithm is used to achieve the best features subset and reduce the complexity ofmodel training. Secondly, the improved ABC algorithm is used to optimize the relatedparameters of kernel ELM model. At last, the well-trained model is used to calculate thepredict results. The experimental results confirm that the valid of the proposed model, andthis hybrid model can not only reduce the features dimension, but also enhance theclassification accuracy. It can be seen that the diagnosis accuracy rate of model is betterthan the existing methods in the area of Thyroid disease diagnosis.(4) Clustering algorithm based kernel ELM model for Parkinson’s disease diagnosisThe redundant and irrelevant features in original feature space can decrease theclassification performance of ELM method. To solve this issue, we combine the optimizdsubtract clustering feature weighting algorithm (ASCFW) with kernel ELM model forParkinson’s disease diagnosis. In this model, ASCFW algorithm is used to transform thefeatures space into linear separable space, differentiate the data that belongs to differentcategories. In addition, ABC algorithm is used adaptively to specify the neighborhoodradius and related parameters of SCFW algorithm, finding the closest cluster centers ofreal data distribution. The experimental results show that the proposed method can not only achieve significantly higher results than the existing methods in terms ofclassification accuracy rates, sensitivity, specificity, computation time and so on, but alsocan make effective diagnosis ofParkinson’s disease.(5) Improved gravitational search algorithm based kernel ELM method for classificationThe related parameters of kernel ELM model and feature selection both can affect theperformance of ELM, however, parameters optimization method and feature selection forELM always perform separately. To solve this problem, we propose a novel adaptiveclassification method based on improve GSA algorithm, IGSA algorithm combines withpattern search to improve the local search ability and convergence speed, and introducesdiversity factor to expand the search space. The discrete and continuous IGSA algorithmsare integrated into a unified algorithm framework, where IGSA is used to optimizefeatures subsets selection and parameters simultaneously. The obtained results show thatthis model can select the most related and important features of credit risk assessment, andthe classification rates are significantly better than the existing and similar methods.更多还原