A MapReduce-Based ELM for Regression in Big Data

Regression is one of the most basic problems in machine learning. In big data era, for regression problem, extreme learning machine (ELM) can get better generalization performance and much fast training speed. However, the enlarging volume of dataset for training makes regression by ELM a challenging task, and it is hard to finish the training in a reasonable time or it will be out of memory. In this paper, through analyzing the theory of ELM, a MapReduce-Based ELM method is proposed. Under the MapReduce framework, ELM submodels are trained in every slave node parallelly. A combination method is designed to combine all the submodels as a complete model. The experiment results demonstrate that the MapReduce-Based ELM can efficient process big dataset on commodity hardware and it has a good performance on speedup under the cloud environment where the dataset is stored as data block in different machines.