An Interval RSP-based ensemble model for big data analysis

Abstract

Ensemble learning for big data has been successful in machine learning and has great advantages over other learning methods. The ensemble model based on Random Sample Partition (RSP) is a prominent method of it. Although the RSP data blocks have the consistent probability distribution function as the whole data, there is some uncertainty in prediction results due to the non-overlapping data between blocks. In this paper, we propose a novel interval ensemble model based on RSP named Inr-RSP, which maps prediction results to interval-valued data by interval modeling and then uses the IAA aggregation method to convert the interval-valued data into fuzzy sets to get a more accurate and stable final result. The experimental classification results from four real datasets also show that the performance of this model is better than that of the traditional RSP ensemble model. And the IAA method usage has a stronger ability to capture uncertainty in prediction than the common majority voting method.