Lagrangian supervised and semi-supervised extreme learning machine

Two extreme learning machine (ELM) frameworks are proposed to handle supervised and semi-supervised classifications. The first is called lagrangian extreme learning machine (LELM), which is based on optimality conditions and dual theory. Then LELM is extended to semi-supervised setting to obtain a semi-supervised extreme learning machine (called Lap-LELM), which incorporates the manifold regularization into LELM to improve performance when insufficient training information is available. In order to avoid the inconvenience caused by matrix inversion, Sherman-Morrison-Woodbury (SMW) identity is used in LELM and Lap-LELM, which leads to two smaller sized unconstrained minimization problems. The proposed models are solvable in a space of dimensionality equal to the number of sample points. The resulting iteration algorithms converge globally and have low computational burden. So as to verify the feasibility and effectiveness of the proposed method, we perform a series of experiments on a synthetic dataset, near-infrared (NIR) spectroscopy datasets and benchmark datasets. Compared with the traditional methods, experimental results demonstrate that the proposed methods achieve better performances than the traditional supervised and semi-supervised methods in most of considered datasets.