Auto associative Extreme Learning Machine based non-linear principal component regression for big data applications

In this paper, we propose a hybrid model that combines the Auto Associative Extreme Learning Machine (AAELM) with Multiple Linear Regression (MLR) (AAELM+MLR) for performing big data regression. It works using Hadoop Mapreduce parallel computing model which is implemented in Python using Dumbo API. It works in two phases. In the first phase, three-layered AAELM is trained. The output of the hidden nodes of AAELM is treated as NLPCs. In the second phase, MLR model is fitted using these NLPCs as input variables. Effectiveness of AAELM+MLR model is demonstrated on two large datasets viz., airline flight delay dataset and gas sensor array dataset, taken from the web. It is observed that AAELM+MLR outperformed MLR model by yielding less average mean squared error (MSE) and MAPE values under the 10 fold cross-validation framework. A statistical test confirms its superiority at 1% level of significance.