A novel approach for extracting ideal exemplars by clustering for massive time-ordered datasets

Abstract

The number and length of massive datasets have increased day by day and this yields more complex machine learning stages due to the high computational costs. To decrease the computational cost many methods were proposed in the literature such as data condensing, feature selection, and filtering. Although clustering methods are generally employed to divide samples into groups, another way of data condensing is by determining ideal exemplars (or prototypes), which can be used instead of the whole dataset. In this study, first the efficiency of traditional data condensing by clustering approach was confirmed according to obtained accuracies and condensing ratios in 9 different synthetic or real batch datasets. This approach was then improved to be employed in time-ordered datasets. In order to validate the proposed approach, 23 different real time-ordered datasets were used in experiments. Achieved mean RMSEs were 0.27 and 0.29 by employing the condensed (mean condensed ratio was 97.17%) and the whole datasets, respectively. Obtained results showed that higher accuracy rates and condensing ratios were achieved by the proposed approach.