Abstract: With the exponentially increasing volume of training data， the performance of centralized ELM with kernels suffers due to large matrix operations. A distributed algorithm named MapReduce based kernelized ELM (MR-KELM) was proposed， which realized an implementation of ELM with kernels on MapReduce in the cloud. The kernel matrix generated by distributed radial basis function was decomposed and then the output weights by distributed multiplication of matrix and vector were calculated by the proposed algorithm. Communications and data exchanges in distributed matrix operations were reduced and good scalability was achieved by MR-KELM. Extensive experiments on synthetic datasets were conducted to verify the training performance and scalability of MR-KELM. Experimental results showed that MR-KELM was effective and efficient for massive learning applications.

Key words: ELM(extreme learning machine), ELM with kernels, classification, distributed, MapReduce