Application of grammar framework to time-series prediction

The previous chapter presented an approach to generate a large number of features using an expert-defined grammar framework. This chapter proceeds to investigate ways to explore such large feature spaces to extract the best features for prediction, i.e. feature selection (FS). Since the proposed framework involves the generation of a large pool of features, there can be redundant and irrelevant features. Therefore, FS is as equally important as feature generation. Several FS and feature extraction techniques can be explored to determine the best approach to discover good feature subsets for particular ML algorithms in different applications. A hybrid feature selection and generation algorithm using grammatical evolution is described as a technique to avoid selective feature pruning by crafting the fitness function to penalise bad feature subsets. The chapter also describes how ML algorithms were used to predict time-series using the sliding window technique, data partitioning, model selection and parameter tuning. © 2015, The Author(s).