A Data-Driven Methodology for Guiding the Selection of Preprocessing Techniques in a Machine Learning Pipeline

The performance of a Machine Learning (ML) model greatly depends on the previous preprocessing of the data. Unfortunately, the decision on which preprocessing techniques should be applied relies on the expertise of data scientists and/or ML practitioners. Since the correct application of some techniques depend on the characteristics of the data whereas others depend on the particular ML model to be trained, this leads to an error-prone process that requires the data scientist to be knowledgeable in all the combinations that may arise. To tackle this problem, we propose a methodology that guides the selection of the most appropriated preprocessing techniques that are highly required or strongly recommended taking into account both the ML model as well as the data characteristics, so that the developer is able to freely experiment with different models while ensuring that no needed preprocessing techniques are overlooked. According to the ML model and the data at hand, the methodology will (i) obtain the characteristics of the model (ii) check whether these characteristics are met by the data or not and (iii) show to the developer which variables require preprocessing and which techniques should be applied so that a proper decision can be made. To the best of our knowledge, this is the only work that tries to gather the most common ML models together with its most adequate preprocessing techniques and encode this information into a methodology that guides this process in a systematic way.