Comparision of diabetic prediction AutoML model with customized model

Automated ML (AutoML) aims to develop Machine Learning (ML) models without much involvement of data scientist. AutoML platforms have automated pre-processing of data such as handling missing values, scaling, dropping duplicates etc and few AutoML platform such as H2O, Data Robot can offer feature engineering. However, all of these AutoML platforms are not capable of integrating domain information into the ML models, which is an exclusive human ability. In this work, the output of AutoML models is compared to the manually created model. Non-invasive diabetes prediction carried out on Pima Indian Diabetes Dataset (PIDD) using TPOT and H2O AutoML platforms which is later compared with the performance of the manual model. Manual model performs outlier detection, data balancing, scaling, hyper parameter tuning and classifies using Adaboost classifier. Highest scores obtained for TPOT model, H2O and manual model are 81.2%, 82% and 100% respectively.