An Accelerated Approach to Parallel Ensemble Techniques Targeting Healthcare and Environmental Applications

Abstract

Ensemble learning techniques adopt comprehensive learning methodologies that produce optimized predictions with reduced variance and bias. The structured Random Forest ensemble learning technique equips a set of weak and diverse decision trees, resulting in an active hybrid learning ensemble. Plagued with high computational complexity, Random Forest Ensemble continues to be the preferred technique when accuracy is of primary importance for learners. Efforts to accelerate the Random Forest Ensembles are in place, however failing to efficiently utilize the data transmission bandwidth between the host and the accelerator hardware. This paper provides an architectural overview of a reconfigurable accelerator based architecture of the Random Forest Ensemble with an efficient data path model for data streaming. The paper also derives the need for an accelerated parallel ensemble method by deriving the results from equivalent sequential software implementations of the algorithm. The validation of the results have been done on healthcare application involving breast cancer classification and environmental applications involving temperature prediction and fuel consumption.