Abstract: Incremental dynamic analysis (IDA) curves contain the uncertainty of seismic input and can reasonably reflect the structural seismic performance. However, its computation process involves a large amount of nonlinear time-history dynamic analyses and thus is computationally expensive. Machine learning (ML) methods have been proven to be a good solution to this problem, but their computational cost is still expensive when the size of training data is large due to the training process involving finding the inverse matrix. To this end, this paper proposes a novel method called low-rank matrix-guided least squares support vector machines for regression (LRLS-SVMR) to overcome these shortcomings. With large-scale training data, LRLS-SVMR is able to build a small-scale low-rank kernel matrix using Nystrom approximation theory for approximating the large-scale original kernel matrix. This allows the training process to only solve the inverse of the small-scale coefficient matrix, which in turn can greatly improve the computational efficiency and maintain high prediction performance. The accuracy and efficiency of the proposed method are verified through a comparison with support vector machines (LS-SVMR) and conventional finite element method (FEM) based on 22,037 seismic response data of reinforced concrete (RC) frames. The results show that the proposed LRLS-SVMR can accurately predict RC frames’ maximum inter-story drift and IDA curves. Its computational cost is nearly 140 times faster than LS-SVMR and 66,000 times faster than FEM.