基于RS-Bag分类器集成技术的上市公司财务危机预测

财务危机预测是金融管理决策中的重要问题,其实质是对未来财务状况的预报和分类。鉴于目前单一分类器预测性能不稳定,本文运用分类器集成技术,以BP神经网络为分类学习算法,建立基于RS-Bag算法的神经网络分类器集成模型。然后,以我国上市公司财务数据为例进行财务危机预测实证研究,结果表明,基于RS-Bag算法的神经网络分类器集成预测精度和泛化性能优于单一神经网络分类器,也优于Bagging分类器集成和RS分类器集成。     

基于MTS特征选择的神经网络集成方法

在Bagging算法基础上,运用马田系统进行特征选择,形成双重扰动改善神经网络集成的分类性能.实验表明,双重扰动增加了集成网络个体精度和差异度,基于MTS-Bagging算法的分类性能相比于Bagging有明显提高.     

一种新的组合模型在信用评分中的应用

针对信用评分的特点,首先对传统的Bagging算法进行改进,然后基于不同的基分类器构建多个改进的Bagging组合模型,最后根据一种全新的方式将其组合,得到双层组合模型.通过实证分析得出以下结论:一、与原Bagging算法相比,改进算法能有效降低第二类错误率;二、与单层组合模型相比,多层组合模型能进一步降低第二类错误率和损失率.     

基于迭代自适应随机决策树的个人信用评估模型研究

信用评估是商业银行控制和防范信贷风险的关键途径.决策树模型较好的直观解释性使其成为当前个人信用评估中的常用模型,但决策树模型存在容易导致过拟合且预测精度有限的问题.通过在决策树模型算法中引入类随机森林随机有放回的抽样模式,运用机器自动循环迭代寻求最优树的建模思想,建立了自适应最优C5.0决策树个人信用评估模型.该模型具有快速收敛特征变量、较好的泛化能力和高预测精度的特点,在实证分析中对商业银行个人信用评估模型质量提升带来比较明显的改进效果.     

集成随机森林和支持向量机的商业银行财务困境预测研究

提出了一种基于随机森林和支持向量机的集成模型来预测商业银行财务困境.结果表明,一方面,与多层感知神经网络相比,支持向量机可以更有效地作为集成学习模型的基分类器,虽然多层感知器神经网络在以往的研究中更多地被用于基分类器.另一方面,与现有的bagging、dagging、multiboost、adaboosting、random subspace等集成学习算法相比,该模型的预测性能明显提高.另一个关键发现是,利用银行业、宏观经济状况和国际金融风险变量补充银行层面的脆弱性,可以显著提高模型在商业银行财务困境预测中的表现.     

基于支持向量机委员会机器的个人信用评估模型

为了充分利用SVM在个人信用评估方面的优点、克服其不足,提出了基于支持向量机委员会机器的个人信用评估模型.将模型与基于属性效用函数估计构造新学习样本方法结合起来进行个人信用评估;经实证分析及与SVM方法对比发现,模型具有更好、更快、更多适应性的预测分类能力.     

基于正态混合模型的贝叶斯分类方法及其应用

本文主要研究正态混合模型的贝叶斯分类方法.贝叶斯分类以后验概率最大为准则,后验概率需要估计相关的条件分布.对于连续型数据的分类,其数据由多个类别混合而成,仅用单一分布难以描述,此时混合模型是一个较好的选择,并且可由EM算法获得.模拟实验表明,基于正态混合模型的贝叶斯分类方法是可行有效的.对于特征较多的分类,不同特征对分类的影响不同,本文对每个特征应用基于正态混合模型的贝叶斯分类方法构建基本分类器,然后结合集成学习,用AdaBoost算法赋予每个分类器权重,再线性组合它们得到最终分类器.通过UCI数据库中实际的Wine Data Set验证表明,本文分类方法与集成学习的结合可以得到高准确率和稳定的分类.     

极限学习机集成在肿瘤分类中的应用

提出一种基于数据集分割的极限学习机集成算法——DS-E-ELM.该算法主要包含以下3个步骤:首先,将数据集分成互不相关的κ个子集,选择κ一1个子集组合成一个训练集,这样可以得到κ个不同的数据集;然后将新得到的κ个数据集利用极限学习机训练得到κ个分类器;最后对κ个分类器预测得到的结果通过多数投票的方法决定预测结果.通过对6个肿瘤数据集的实验证明,DS-E-ELM与单独的ELM、Bagging、Boosting等算法相比,具有更高的分类精度,且稳定性更好.     

基于BalanceCascade-GBDT算法的类别不平衡虚假评论识别方法

虚假评论是电商发展过程中一个无法避免的难题. 针对在线评论数据中样本类别不平衡情况,提出基于BalanceCascade-GBDT算法的虚假评论识别方法. BalanceCascade算法通过设置分类器的误报率逐步缩小大类样本空间,然后集成所有基分类器构建最终分类器. GBDT以其高准确性和可解释性被广泛应用于分类问题中,并且作为样本扰动不稳定算法,是十分合适的基分类模型. 模型基于Yelp评论数据集,采用AUC值作为评价指标,并与逻辑回归、随机森林以及神经网络算法进行对比,实验证明了该方法的有效性.     

基于集成算法的股票指数预测

利用集成算法中的Bagging、Boosting和Random Forest三个方法,选取股票指数中的中小板指数、深证成指数、上证指数、创业板指数4组数据进行分析,得出Random Forest对上证指数、中小板指预测结果较好;Boosting对创业板指预测结果较好;Bagging对深证成指预测较好.并在4个板指中,随机选取了4支股票数据(分别为大连重工、中南建设、中国医药、东方国信)进行分析,得出集成算法在数据为200个的情况下,预测结果较为准确,其中不同方法对不同股票的适宜程度有所不同.     

Qualitative Analysis of Bifurcating Solutions in the Lengyel-Epstein Model

One of the most fundamental problems in theoretical biology is to explain the mechanisms by which patterns and forms are created in the＇living world. In his seminal paper ＂The Chemical Basis of Morphogenesis＂, Turing showed that a system of coupled reaction-diffusion equations can be used to describe patterns and forms in biological systems. However, the first experimental evidence to the Turing patterns was observed by De Kepper and her associates（1990） on the CIMA reaction in an open unstirred reactor, almost 40 years after Turing＇s prediction. Lengyel and Epstein characterized this famous experiment using a system of reaction-diffusion equations. The Lengyel-Epstein model is in the form as follows     

(0,1 ;0)-INTERPOLATION ON INFINITE INTERVAL (-∞, +∞)

In this paper, we study the explicit representation and convergence of (0, 1; 0)-interpolation on infinite interval, which means to determine a polynomial of degree ≤ 3n - 2 when the function values are prescribed at two set of points namely the zeros of Hn(x) and H′n(x) and the first derivatives at the zeros of H′n(x).     

RECENT PROGRESS ON SPHERICAL HARMONIC APPROXIMATION MADE BY BNU RESEARCH GROUP ——In memory of Professor Sun Yongsheng

As early as in 1990, Professor Sun Yongsheng, suggested his students at Beijing Normal University to consider research problems on the unit sphere. Under his guidance and encouragement his students started the research on spherical harmonic analysis and approximation. In this paper, we incompletely introduce the main achievements in this area obtained by our group and relative researchers during recent 5 years （2001-2005）. The main topics are： convergence of Cesaro summability, a.e. and strong summability of Fourier-Laplace series; smoothness and K-functionals; Kolmogorov and linear widths.     

A REPRESENTATION FORMULA RELATED TO SCHR(O)DINGER OPERATORS

Schr(o)dinger operator is a central subject in the mathematical study of quantum mechanics.Consider the Schrodinger operator H = -△ V on R, where △ = d2/dx2 and the potential function V is real valued. In Fourier analysis, it is well-known that a square integrable function admits an expansion with exponentials as eigenfunctions of -△. A natural conjecture is that an L2 function admits a similar expansion in terms of "eigenfunctions" of H, a perturbation of the Laplacian (see [7], Ch. Ⅺ and the notes), under certain condition on V.     

Journal of the Operations Research Society of China Special issue on Operations Research in Healthcare Management

正Guest Editors:Hong Chen,Shanghai Jiao Tong University,Shanghai,China Guohua Wan,Shanghai Jiao Tong University,Shanghai,China David Yao,Columbia University,New York,USA Scope:Healthcare delivery worldwide has been fraught with high cost,low efficiency and poor quality of patient care service.For the field of operations research(OR),healthcare offers some of the biggest challenges as well as best opportunities in     

ANALYSIS IN THEORY AND APPLICATIONS Vol. 30, 2014 CONTENTS

正~~     

An alternative mathematical algorithm for the photo- and videokeratoscope

Due to the resolution of current laser technology, the accuracy of corneal topography as measured by the videokeratoscope is no longer adequate to provide precise enough data for refractive surgery or for the fitting of customized contact lenses. We present an algorithm for recovering corneal topography that makes use of modern differential geometric techniques and numerical descent in Sobolev spaces. We believe this algorithm may be used with the photo- and videokeratoscope to increase the accuracy of the recovered corneal topography.     

A note on Sobolev orthogonality for Laguerre matrix polynomials

Let {Ln(A,λ)(x)}n≥0 be the sequence of monic Laguerre matrix polynomials defined on [0, ∞) by Ln(A,λ)(x)=n!/(-λ)n∑nk=0(-λ)κ/k!(n-1)! (A I)n[(A I)k]-1 xk,where A ∈ Cr×r. It is known that {Ln(A,λ)(x)}n≥0 is orthogonal with respect to a matrix moment functional when A satisfies the spectral condition that Re(z) ＞ - 1 for every z ∈σ(A).In this note we show that forA such that σ(A) does not contain negative integers, the Laguerre matrix polynomials Ln(A,λ) (x) are orthogonal with respect to a non-diagonal SobolevLaguerre matrix moment functional, which extends two cases: the above matrix case and the known scalar case.     

CONTENTS OF VOLUME 30

正~~