Abstract | Current advances in communication, sensor and computing technologies are generating information in never before seen amounts and at constantly increasing rates (i.e. the information explosion, the Internet of Things). From the point of view of data analytics, the information is composed of a diversity of data types and it contains uncertainties and incompleteness of different degrees, which add an extra component to the original heterogeneity. Many data mining and machine learning methods do not handle heterogeneity well. Extreme learning machines (ELM) are interesting computational algorithms because of their simplicity, their good performance and their speed. They can be extended for processing information composed of heterogeneous data types (HT-ELM), capable of addressing classification and regression problems with complex data. Two approaches are discussed: one works directly with the heterogeneous data and the other one transforms the information into simpler homogeneous spaces that preserve structural properties. In them, standard learning methods can be applied, including classical ELMs among others. Both approaches are illustrated using real world examples involving heterogeneous predictor variables composed of mixtures of nominal, ordinal, interval, ratio, fuzzy variables, and entire empirical probability distributions. In all cases HT-ELM and ELM models produced results that compare favorably with well-established methods. |
---|