A robust extreme learning machine framework for uncertain data classification

Uncertain or missing data may occur in many practical applications. A principled strategy for handling this problem would therefore be very useful. We consider two-class and multi-class classification problems where the mean and covariance of each class are assumed to be known. With simple structure, fast speed and good performance, extreme learning machine (ELM) has been an important technology in machine learning. In this work, from the viewpoint of probability, we present a robust ELM framework (RELM) for missing data classification. Applying the Chebyshev–Cantelli inequality, the proposed RELM is reformulated as a second-order cone programming with global optimal solution. The proposed RELM only relates to the second moments of input samples and makes no assumption about the data probability distribution. Expectation maximization algorithm is used to fill in missing values and then obtain complete data. Numerical experiments are simulated in various datasets from UCI database and a practical application database. Experimental results show that the proposed method can achieve better performance than traditional methods. These results illustrate the feasibility and effectiveness of the proposed method for missing data classification.