Optimization of Stacked Unsupervised Extreme Learning Machine to improve classifier performance

In the era of Big Data, data size and data security are issues that need to be solved. To address this problem, we may apply data compression technique or data encryption. On the other hand, the coding based method may be other solution. The coding based method may learn data distribution and reduce data dimension with minimized the loss of information from the original data. Stacked Unsupervised Extreme Learning Machine (Stacked US-ELM) is one of the fastest methods which can be used to address it. The problem is how many stacks we need to get the optimal performance of the classifier. In this research, we inspected the performance of Stacked Unsupervised Extreme Learning Machine (US-ELM) for enhancing classifier performance and proposed a new method. The proposed method is a loop scheme of Stacked US-ELM to optimize the number of stacks of US-ELM. We conducted the experiment using ECG-sleep dataset, synthetic dataset, and Glass dataset. To measure the performance of Stacked US-ELM and the proposed method, we conducted classification according to our data sets using Support Vector Machine (SVM). The 5-Folds Cross Validation is used to evaluate the performance. We compared the result of the classification without US-ELM with the fix Stacked US-ELM and the proposed method. The results showed the proposed method achieved the best performance over Stacked US-ELM in all dataset. The mean of accuracies of the proposed method are 64,39%, 76,24%, 63,22%, and 69, 61% on 4 class ECG-sleep dataset, 3 class ECG-sleep dataset, skewed synthetic dataset, and glass dataset.