Résumé
As illustrated in Fig. 1, data cleaning and preparation require a sophisticated sequence of tasks for the detection and elimination of a variety of intricate data quality problems (e.g., duplicates, inconsistent, missing, and outlying values). Generally, the users may not know which preprocessing methods can be applied to optimize the final results downstream. This would require executing all possible methods for each task of preprocessing, as well as all the possible combinations of the methods with different orderings and configurations. These methods can be applied to the whole or some parts of the dataset with eventual re-iterations. AutoML approaches can optimize the hyper-parameters of a considered ML model, but they support only a limited number of preprocessing steps with bydefault methods. Recently, Alpine Meadow [3] combines an Au-toML approach and a cost model to select candidate logical ML pipeline plans (as in DB query optimization). Multi-armed bandits are used to select promising logical ML pipeline plans, and Bayesian Optimization is used to fine-tune the hyper-parameters of the selected models in the search space. Although Alpine Meadow proposes an Adaptive Pipeline Selection (APS) method to find a trade-off between speed and accuracy of ML pipeline plan evaluation and returns the results progressively, the human expertise is not actually exploited for building or pruning the data preparation pipeline search space, and the user has a passive role with limited interaction. We argue that data cleaning and preparation are intrinsically "AI-hard" as they can hardly be achieved by a fully automated system, and more efforts should be devoted to proposing a principled and efficient data preparation approach to help and learn from the user in selecting the optimal sequence of data preparation tasks. Improving the quality of input data for ML and leveraging human expertise, subsequent learning performance will benefit.