Matching Items (5)
Filtering by
- All Subjects: Data Mining
- Creators: Montgomery, Douglas C.
Description
This work presents two complementary studies that propose heuristic methods to capture characteristics of data using the ensemble learning method of random forest. The first study is motivated by the problem in education of determining teacher effectiveness in student achievement. Value-added models (VAMs), constructed as linear mixed models, use students’ test scores as outcome variables and teachers’ contributions as random effects to ascribe changes in student performance to the teachers who have taught them. The VAMs teacher score is the empirical best linear unbiased predictor (EBLUP). This approach is limited by the adequacy of the assumed model specification with respect to the unknown underlying model. In that regard, this study proposes alternative ways to rank teacher effects that are not dependent on a given model by introducing two variable importance measures (VIMs), the node-proportion and the covariate-proportion. These VIMs are novel because they take into account the final configuration of the terminal nodes in the constitutive trees in a random forest. In a simulation study, under a variety of conditions, true rankings of teacher effects are compared with estimated rankings obtained using three sources: the newly proposed VIMs, existing VIMs, and EBLUPs from the assumed linear model specification. The newly proposed VIMs outperform all others in various scenarios where the model was misspecified. The second study develops two novel interaction measures. These measures could be used within but are not restricted to the VAM framework. The distribution-based measure is constructed to identify interactions in a general setting where a model specification is not assumed in advance. In turn, the mean-based measure is built to estimate interactions when the model specification is assumed to be linear. Both measures are unique in their construction; they take into account not only the outcome values, but also the internal structure of the trees in a random forest. In a separate simulation study, under a variety of conditions, the proposed measures are found to identify and estimate second-order interactions.
ContributorsValdivia, Arturo (Author) / Eubank, Randall (Thesis advisor) / Young, Dennis (Committee member) / Reiser, Mark R. (Committee member) / Kao, Ming-Hung (Committee member) / Broatch, Jennifer (Committee member) / Arizona State University (Publisher)
Created2013
Description
Random Forests is a statistical learning method which has been proposed for propensity score estimation models that involve complex interactions, nonlinear relationships, or both of the covariates. In this dissertation I conducted a simulation study to examine the effects of three Random Forests model specifications in propensity score analysis. The results suggested that, depending on the nature of data, optimal specification of (1) decision rules to select the covariate and its split value in a Classification Tree, (2) the number of covariates randomly sampled for selection, and (3) methods of estimating Random Forests propensity scores could potentially produce an unbiased average treatment effect estimate after propensity scores weighting by the odds adjustment. Compared to the logistic regression estimation model using the true propensity score model, Random Forests had an additional advantage in producing unbiased estimated standard error and correct statistical inference of the average treatment effect. The relationship between the balance on the covariates' means and the bias of average treatment effect estimate was examined both within and between conditions of the simulation. Within conditions, across repeated samples there was no noticeable correlation between the covariates' mean differences and the magnitude of bias of average treatment effect estimate for the covariates that were imbalanced before adjustment. Between conditions, small mean differences of covariates after propensity score adjustment were not sensitive enough to identify the optimal Random Forests model specification for propensity score analysis.
ContributorsCham, Hei Ning (Author) / Tein, Jenn-Yun (Thesis advisor) / Enders, Stephen G (Thesis advisor) / Enders, Craig K. (Committee member) / Mackinnon, David P (Committee member) / Arizona State University (Publisher)
Created2013
Description
Supply chains are increasingly complex as companies branch out into newer products and markets. In many cases, multiple products with moderate differences in performance and price compete for the same unit of demand. Simultaneous occurrences of multiple scenarios (competitive, disruptive, regulatory, economic, etc.), coupled with business decisions (pricing, product introduction, etc.) can drastically change demand structures within a short period of time. Furthermore, product obsolescence and cannibalization are real concerns due to short product life cycles. Analytical tools that can handle this complexity are important to quantify the impact of business scenarios/decisions on supply chain performance. Traditional analysis methods struggle in this environment of large, complex datasets with hundreds of features becoming the norm in supply chains. We present an empirical analysis framework termed Scenario Trees that provides a novel representation for impulse and delayed scenario events and a direction for modeling multivariate constrained responses. Amongst potential learners, supervised learners and feature extraction strategies based on tree-based ensembles are employed to extract the most impactful scenarios and predict their outcome on metrics at different product hierarchies. These models are able to provide accurate predictions in modeling environments characterized by incomplete datasets due to product substitution, missing values, outliers, redundant features, mixed variables and nonlinear interaction effects. Graphical model summaries are generated to aid model understanding. Models in complex environments benefit from feature selection methods that extract non-redundant feature subsets from the data. Additional model simplification can be achieved by extracting specific levels/values that contribute to variable importance. We propose and evaluate new analytical methods to address this problem of feature value selection and study their comparative performance using simulated datasets. We show that supply chain surveillance can be structured as a feature value selection problem. For situations such as new product introduction, a bottom-up approach to scenario analysis is designed using an agent-based simulation and data mining framework. This simulation engine envelopes utility theory, discrete choice models and diffusion theory and acts as a test bed for enacting different business scenarios. We demonstrate the use of machine learning algorithms to analyze scenarios and generate graphical summaries to aid decision making.
ContributorsShinde, Amit (Author) / Runger, George C. (Thesis advisor) / Montgomery, Douglas C. (Committee member) / Villalobos, Rene (Committee member) / Janakiram, Mani (Committee member) / Arizona State University (Publisher)
Created2012
Description
Public health surveillance is a special case of the general problem where counts (or rates) of events are monitored for changes. Modern data complements event counts with many additional measurements (such as geographic, demographic, and others) that comprise high-dimensional covariates. This leads to an important challenge to detect a change that only occurs within a region, initially unspecified, defined by these covariates. Current methods are typically limited to spatial and/or temporal covariate information and often fail to use all the information available in modern data that can be paramount in unveiling these subtle changes. Additional complexities associated with modern health data that are often not accounted for by traditional methods include: covariates of mixed type, missing values, and high-order interactions among covariates. This work proposes a transform of public health surveillance to supervised learning, so that an appropriate learner can inherently address all the complexities described previously. At the same time, quantitative measures from the learner can be used to define signal criteria to detect changes in rates of events. A Feature Selection (FS) method is used to identify covariates that contribute to a model and to generate a signal. A measure of statistical significance is included to control false alarms. An alternative Percentile method identifies the specific cases that lead to changes using class probability estimates from tree-based ensembles. This second method is intended to be less computationally intensive and significantly simpler to implement. Finally, a third method labeled Rule-Based Feature Value Selection (RBFVS) is proposed for identifying the specific regions in high-dimensional space where the changes are occurring. Results on simulated examples are used to compare the FS method and the Percentile method. Note this work emphasizes the application of the proposed methods on public health surveillance. Nonetheless, these methods can easily be extended to a variety of applications where counts (or rates) of events are monitored for changes. Such problems commonly occur in domains such as manufacturing, economics, environmental systems, engineering, as well as in public health.
ContributorsDavila, Saylisse (Author) / Runger, George C. (Thesis advisor) / Montgomery, Douglas C. (Committee member) / Young, Dennis (Committee member) / Gel, Esma (Committee member) / Arizona State University (Publisher)
Created2010
Description
The Partition of Variance (POV) method is a simplistic way to identify large sources of variation in manufacturing systems. This method identifies the variance by estimating the variance of the means (between variance) and the means of the variance (within variance). The project shows that the method correctly identifies the variance source when compared to the ANOVA method. Although the variance estimators deteriorate when varying degrees of non-normality is introduced through simulation; however, the POV method is shown to be a more stable measure of variance in the aggregate. The POV method also provides non-negative, stable estimates for interaction when compared to the ANOVA method. The POV method is shown to be more stable, particularly in low sample size situations. Based on these findings, it is suggested that the POV is not a replacement for more complex analysis methods, but rather, a supplement to them. POV is ideal for preliminary analysis due to the ease of implementation, the simplicity of interpretation, and the lack of dependency on statistical analysis packages or statistical knowledge.
ContributorsLittle, David John (Author) / Borror, Connie (Thesis advisor) / Montgomery, Douglas C. (Committee member) / Broatch, Jennifer (Committee member) / Arizona State University (Publisher)
Created2015