Grammar-Based Evolutionary Approach for Automatic Workflow Composition with Open Preprocessing Sequence

Knowledge discovery is a complex process involving several phases. Some of them are repetitive and time-consuming, so they are susceptible of being automated. As an example, the large number of machine learning algorithms, together with their hyper-parameters, constitutes a vast search space to explore. In this vein, the term AutoML was coined to encompass those approaches automating such phases. The automatic workflow composition is an AutoML task that involves both the selection and the hyper-parameter optimisation of the algorithms addressing different phases, thus giving a more comprehensive assistance during the knowledge discovery process. Unlike other proposals that predetermine the structure of the preprocessing sequence, and in some cases the size of the workflow, our proposal generates workflows made up of an arbitrary number of preprocessing algorithms of any type and a classifier. This allows returning more accurate results since its avoids the oversimplification of the solution space. The optimisation is conducted by a grammar-guided genetic programming algorithm. The proposal has been validated and compared against TPOT and RECIPE generating workflows with greater predictive performance.