Transcript DISSERTATION PROPOSAL DEFENSE

Boosting and Bagging
For Fun and Profit
Hal Elkins
David Lucus
Keith Walker
Ensemble Methods
 Improve predictive performance of a given statistical model





fitting technique.
Run a base procedure many times while changing input data
Estimates are linear or non-linear combinations of iteration
estimates
Originally used for machine learning and data and text
mining
Attracting attention due to relative simplicity and popularity
of bootstrapping
Are ensemble methods useful to academic researchers?
Bagging
 Bootstrap aggregating for improving unstable estimation






schemes - Breiman (1996)
Variance reduction for base procedure – Bühlmann and Yu
(2002)
Bagging requires user specified input models
Step 1) construct bootstrap sample with replacement.
Step 2) compute estimator
Step 3) repeat
Base procedure bias is increased
Boosting
 Boosting proposed by Schapire (1990) and Freund (1995)
 Nonparametric optimization - useful if we have no idea for a





model
Bias reduction
Step 1) initialize – apply base procedure
Step 2) compare residuals
Step 3) repeat
Variance is increased
Enterprise Miner: Bagging & Boosting
Enterprise Miner: Bagging
 Control:
 Ensemble (under model menu)
 Inputs
 The outputs of other models
 Regression
 Decision Tree
 Neural Networks
 Settings
 Limited and VERY BLACK BOX
Enterprise Miner: Bagging
 How to view output
 Connect Ensemble output to Regression node
 Use Comparison node to:
 Compare Bagged model with input models
 Note:
 Ensemble will only out perform the input models IF, there is large
disagreement in the input models. (AAEM_61 Manual)
Year 1 Analysis
Year 2 Analysis
Enterprise Miner: Boosting
 Control
 Gradient Boosting (under model menu)
 Input
 Data Partition or Dataset
 Settings (MANY)
 Assessment Measures
 Tree Size Settings
 Iterations
 Etc.
Enterprise Miner: Boosting
 Outputs
 List variable importance
 List # of decision rules containing each variable
 Hook to regression node for more information
 Compare with other models
 Use Model Comparison node
 Example
 Gradient Boost Regression AIC: -5868.21
 Base Regression AIC: -2866.43
 Base Neural Network AIC: -2981.86
Enterprise Miner: Boosting
 A Boosting Story (Another data set)
 Prediction of graduation from TTU
 Data 2004-2007 (SAT, ACT, HSRank%, Parent Education,
Income Level)
 Texas Census Level (Matched on Student’s High School
county code (20+ variables)
 After boosting 1 variable had prediction power of graduation
from TTU
Previous Use
 Manescu, C. & Starica, C. 2009. Do corporate social
responsibility scores explain and predict firm
profitability? A case study on the publishers of the Dow
Jones Sustainability Indexes. Working Paper, Gothenburg
University.
 Bagging and Boosting to determine if CSR measures
affect ROA
Model Comparisons – Data Courtesy Dr. Romi
 Dependent Variable (Change in CSR Performance)i,t+n
 OLS = α + β1[CSO]i,t + β2COMMITTEEi,t + β3∆SIZEi,t+1 +
β4∆ROAi,t+1 + β5ΔFINi,t+1 + β6ΔLEVi,t+1 + β7GLOBALi,t +
β8CEOCHAIRi,t + β9HIERi,t + β10ESIi,t + β11LITIGATIONi,t +
β12EXPERTi,t + ε
 Boosting = α + β1[CSO]i,5 + β2COMMITTEEi,t +
β3GLOBALi,t + β4CEOCHAIRi,t + β5HIERi,t + ε
 Bagging All = α + β1[CSO]i,4 + β2[CSO]i,5 + β3COMMITTEEi,t
+ β4GLOBALi,t + β5EXPERTi,t + ε
Is Either Useful to Us?
 What we thought
 Both useful in model selection and refinement
 What we concluded
 Bagging for settling different model possibilities
 Bagging helps determine model disagreement on “black box”
models
 Boosting for grounded theory
 Boosting for starting model point
References
 Breiman, L.: Bagging predictors. Mach. Learn. 24, 123–140 (1996a)
 Bühlmann, P.,Yu, B.: Analyzing bagging. Ann. Stat. 30, 927–961 (2002)
 Freund, Y.: Boosting a weak learning algorithm by majority. Inform.
Comput. 121, 256–285 (1995)
 Schapire, R.E.: The strength of weak learnability. Mach. Learn. 5, 197–
227 (1990)
 Course Notes
 George, Jim et al: Applied Analytics Using SAS Enterprise Miner 6.1,
Course Notes, 2009
 Working Paper
 Manescu, C. & Starica, C. 2009. Do corporate social responsibility scores
explain and predict firm profitability? A case study on the publishers of the
Dow Jones Sustainability Indexes. Working Paper, Gothenburg University.
THANK YOU!
 QUESTIONS?