Nowadays, the Extreme learning machine (ELM) is playing a key role in machine intelligence and big data analytics due to its various advantages such as fast training rate, universal classification/regression and the capability of approximation. The standard ELM uses the Moore-Penrose generalized pseudo-inverse for solving the hidden layer activation matrix and also it identifies the output weights. Because of that, the standard ELM takes more time to train the features from the dataset. In ELM, scalability also considered as a one of the major concern while processing the large dataset. In order to overcome this concern, the Automotive Rank based ELM (AR-ELM) is proposed to obtain an effective tensor decomposition for diminishing the training time. Besides, the Bayesian approach is considered in this AR-ELM to remove the redundancy from the decomposed samples of the tensor. The major objective of this proposed AR-ELM is to process the large amount of dataset without depending on memory capacities. The recognition accuracy is improved by eliminating redundant information. The key idea of the ARELM is to reduce the training time while processing the huge dataset. The implementation and simulation of the ARELM is done in Spark Python 3.7. The performance of the AR-ELM is analysed in terms of accuracy, precision, recall and training time. The proposed methodology is compared with three existing methodologies such as basic ELM, ELM-TUCKER and ELM-PARAFAC. The recognition accuracy of the AR-ELM methodology with Hardlims activation function is 0.8879 for letter recognition dataset, it is high when compared to the basic ELM, ELM-TUCKER and ELM-PARAFAC that are 0.8102, 0.8375 and 0.834 respectively. © 2019 Intelligent Network and Systems Society.