Constructive model selection for multi-output extreme learning machine

In this paper, a novel constructive model selection for multi-output extreme learning machine (CMS-MELM) is proposed to deal with multi-output regressions. The significant contributions to this paper feature the key characteristics as follows. 1) The initial candidate pool for CMS-MELM is randomly generated according to the ELM strategy, and ranked chunk-by-chunk based on a novel improved multi-response sparse regression (I-MRSR) incorporated with λ weighting. 2) Accordingly, the proposed constructive model selection works with fast speed due to chunk-type training process, which also benefits stable hidden node selection and corresponding generalization capability. 3) Furthermore, validation and retraining phases are conducted to enhance the overall performance of the resulting CMS-MELM scheme. Finally, the convincing performance of the complete CMS-MELM paradigm is verified by simulation studies on real-life benchmark multi-output regressions. Comprehensive comparisons of the CMS-MELM with other well-known strategies, i.e., ELM and OP-ELM, indicate the remarkable superiority in terms of generalization capability and stable compact structure. Clear conclusions are steadily drawn that the CMS-MELM method is feasibly effective for multi-output regressions.