Heterogeneous extreme learning machines

Abstract

The developments in communication, sensor and computing technologies are generating information at increasing rates and the nature of the data is becoming highly heterogeneous. Accordingly, the objects under study are described by collections of variables of very different kinds (e.g. numeric, non-numeric, images, signals, videos, documents, etc.) with different degrees of imprecision and incompleteness. Many data mining and machine learning methods do not handle heterogeneity well, requiring variables of the same type, information completeness (or imputation), also assuming no imprecision. Extreme learning machines (ELM) are very interesting computational algorithms because of their structural simplicity, their good performance and their speed. Accordingly, extending their scope by making them capable of processing heterogeneous information may increase their attractiveness as a modeling tool for addressing complex problems. ELMs are discussed in the context of heterogeneous data and approaches to build ELMs capable of performing classification and regression tasks in such cases are presented. Their performance is illustrated with real world examples of classification and regression involving heterogeneous information with scalar data described by nominal, ordinal, interval, ratio, and fuzzy variables as well as with entire empirical probability distributions as predictor variables.