CS539 Project Report -- Evaluating hypothesis Mingyu Feng Feb 24th, 2004
Learning J48 Decision Tree Dataset: S1, 45 nominal and 10 real attributes. No preprocessing Result Correctly Classified Instances 633 63.3 % Incorrectly Classified Instances 367 36.7 % Kappa statistic 0.403 Mean absolute error 0.1177 Root mean squared error 0.2946 Relative absolute error 65.7543 % Root relative squared error 98.6013 % errorS1(t) = 0.367 95% confidence interval = 0.367 +/- 0.2987. Feb 24th, 2005
Learning Neural Networks Dataset: S1, 45 nominal and 10 real attributes. No preprocessing Result Correctly Classified Instances 648 64.8 % Incorrectly Classified Instances 352 35.2 % Kappa statistic 0.4266 Mean absolute error 0.1075 Root mean squared error 0.2888 Relative absolute error 60.1082 % Root relative squared error 96.6877 % errors2(nn) = 0.352 Feb 24th, 2005
Difference between True Errors errorS1(t) = 0.367, errors2(nn) = 0.352 d = 0.015, σ = 0.0215 95% two-sided confidence interval for difference between two true errors 0.015 +/- 0.042 Feb 24th, 2005
Compare Learning Algorithms k =11 1100 instances in D0 divided into 11 partitions (T1… T11), 100 instances each For i =1 to 11, use Ti as test set, train decision tree and neural networks on remaining data Expectation of i =-0.0082, standard derivation = 0.00784 95% confidence interval for estimating the difference in error between J4.8 decision trees and neural networks: -0.0082 +/- 0.0175 i 1 2 3 4 5 6 7 8 9 10 11 error Ti (t) 0.37 0.28 0.39 0.29 0.32 0.33 error Ti (nn) 0.40 0.34 0.31 0.38 0.3 0.36 0.35 i 0.04 -0.03 0.03 0.01 -0.01 -0.04 -0.02 Feb 24th, 2005