Evolving on-line prediction model dealing with industrial data sets

In this work we present an instance of an architecture for the development of robust evolving predictive models. The architecture provides a conceptual framework for the development of such models while at the same time it provides mechanisms for the minimisation of effort needed for the development and maintenance of the models. These mechanisms deal with the model and parameter selection, model training, validation and adaptation. Another challenge for the proposed instance is to deal with an industrial data set containing several issues like missing data, outliers, drifting data, etc. This fact calls for high robustness of the deployed models. The success of the models lays in the goal oriented application of several concepts like ensemble building, local learning, parameter cross-validation which are provided by the architecture and exploited by the discussed instance.