A boosting method for metamodel generation based on large data

Abstract

The purpose of this study is to solve a problem which is impossibilityof handling large data, common for conventional metamodel generationtechnique. In this study, ensemble model method is used, especiallyboosting, which is used to informationize of big data in the field of industrial engineering. This study approaches the problem by the method of using errors of regression sample, not using weight update criterion through binarization of regression sample which is characterized by regression model generation technique that uses boosting. This study changes error decision criterion which is used in boosting by using a concept of online learning. This study proposes algorithm that enables regression model generation by using large data. Proposed algorithm is evaluated by six mathematical examples. As algorithm for benchmarking, bagging for regression and AdaBoost for regression which is modification of AdaBoost for classification are used. Predicted response value is calculated by using each algorithm for 1000 test points which is sampled in advance as the measure of comparison of performance. Accuracy is evaluated through average and standard deviation of errors, between calculated and real response value. Efficiency is evaluated through times of generating each model which is required to generate one ensemble model. With respect to average of error, proposed algorithm improves by 0%~88% on average 41.5% against bagging. With respect to standard deviation of error, proposed algorithm improves by 0%~97%, on average 53.8% against bagging. With respect to efficiency, proposed algorithm improves by 1.08~2.11 times, on average 1.45 times against bagging. With respect to average of Error, proposed algorithm improves by 0%~58.6%, on average 21.4% against AdaBoost for regression. With respect to standard deviation of error, proposed algorithm improves by 0%~72.7%, on average 25.4% against AdaBoost for regression. With respect to efficiency, proposed algorithm improves by 0%~96.4%, on average 70.3% against AdaBoost for regression. Findings of this study suggest that proposed technique in this study is superior to previous techniques.