Search

The Sure Independence Screening (SIS) provides a fast and efficient ranking for the importance of variables for ultra-high dimensional regressions. However, classical SIS cannot eliminate false importance in the ranking, which is exacerbated in nonparametric settings. To address this problem, a novel screening approach is proposed by partitioning the sample into subsets sequentially and creating a tree-like structure of sub-samples called SIS-tree. SIS-tree is straightforward to implement and can be integrated with various measures of dependence. Theoretical results are established to support this approach, including its "sure screening property". Additionally, SIS-tree is extended to a forest with improved performance. Through simulations, the proposed methods are demonstrated to have great improvement comparing with existing SIS methods. The selection of a cutoff for the screening is also investigated through theoretical justification and experimental study. As a direct application, classifications of high-dimensional data are considered, and it is found that the screening and cutoff can substantially improve the performance of existing classifiers. The proposed approaches can be implemented using R package "SIStree" at https://github.com/liuyu-star/SIStree. • A model-free variable selection method with tree and forest structures is introduced with any dependence measure. • The proposed methods possess the sure screening property and can mitigate selecting unimportant variables. • A cross-validation procedure is proposed to determine the optimal cutoff for classifications and regressions. • The new method is compared with other variable selections through a variety of simulations and real data examples. [ABSTRACT FROM AUTHOR]

Abstract: Predicting business failure is an important and challenging issue that has served as an impetus for many academic studies over the past three decades. This study aims at developing CBR-based hybrid models of predicting business failure. The need to supplement CBR (Case-Based Reasoning) with other classification and diagnosis techniques is triggered by the fact that accuracy and effectiveness tend to get reduced when CBR alone is applied to handle too many attributes. To enhance the accuracy of bankruptcy prediction, the hybrid models developed by this study include: RST–CBR (combining Rough Set Theory with CBR), RST–GRA–CBR (integrating RST, Grey Relational Analysis, and CBR), and CART–CBR (combining Classification and Regression Tree with CBR). In order to verify the ability of the proposed models to achieve optimal accuracy rate, this study further compares the predictive ability of CBR with those of other comparative models. [Copyright &y& Elsevier]

Abstract: A variable selection method for constructing decision trees with rank data is proposed. It utilizes conditional independence tests based on loglinear models for contingency tables. Compared with other selection methods, our method is computationally more efficient. Moreover, our method is relatively unbiased and powerful in selecting the correct split variables. Simulation results and a real data study are given to demonstrate the strength of our method. [Copyright &y& Elsevier]

Abstract: We propose a new nonlinear classification method based on a Bayesian “sum-of-trees” model, the Bayesian Additive Classification Tree (BACT), which extends the Bayesian Additive Regression Tree (BART) method into the classification context. Like BART, the BACT is a Bayesian nonparametric additive model specified by a prior and a likelihood in which the additive components are trees, and it is fitted by an iterative MCMC algorithm. Each of the trees learns a different part of the underlying function relating the dependent variable to the input variables, but the sum of the trees offers a flexible and robust model. Through several benchmark examples, we show that the BACT shows excellent performance. We apply the BACT technique to classify whether firms would be insolvent. This practical example is very important for banks to construct their risk profile and operate successfully. We use the German Creditreform database and classify the solvency status of German firms based on financial statement information. We show that the BACT is a serious competitor to the logit model, CART, the Support Vector Machine, random forest and gradient boosting. [Copyright &y& Elsevier]

Abstract: Credit scoring has become a very important task as the credit industry has been experiencing severe competition during the past few years. The artificial neural network is becoming a very popular alternative in credit scoring models due to its associated memory characteristic and generalization capability. However, the relative importance of potential input variables, long training process, and interpretative difficulties have often been criticized and hence limited its application in handling credit scoring problems. The objective of the proposed study is to explore the performance of credit scoring using two commonly discussed data mining techniques—classification and regression tree (CART) and multivariate adaptive regression splines (MARS). To demonstrate the effectiveness of credit scoring using CART and MARS, credit scoring tasks are performed on one bank credit card data set. As the results reveal, CART and MARS outperform traditional discriminant analysis, logistic regression, neural networks, and support vector machine (SVM) approaches in terms of credit scoring accuracy and hence provide efficient alternatives in implementing credit scoring tasks. [Copyright &y& Elsevier]

Inferring causal effect from observational studies is a central topic in many scientific fields, including social science, health and medicine. The statistical methodology for estimating population average causal effect has been well established. However, the methods for identifying and estimating subpopulation causal effects are relatively less developed. Part of the challenge is that the subgroup structure is usually unknown, therefore, methods working well for population level effect need to be modified to address this. A tree method based on a matched design is proposed to identify subgroups with differential treatment effects. To remove observed confounding, propensity-score-matched pairs are first created. Then the classification and regression tree is applied to the within-pair outcome differences to identify the subgroup structure. This nonparametric approach is robust to model misspecification, which is important because it becomes much harder to specify a parametric outcome model in the presence of subgroup effects. In addition to describing assumptions under which our matching estimator is unbiased, algorithms for identifying subgroup structures are provided. Simulation results indicate that the proposed approach compares favorably, in terms of the percentage of correctly identifying true tree structure, with other competing tree-based methods, including causal trees, causal inference trees and the virtual twins approach. Finally the proposed method is implemented to examine the potential subgroup effect of the timing of Tobramycin use on chronic infection among pediatric Cystic Fibrosis patients. [ABSTRACT FROM AUTHOR]