Automating data preparation pipeline efficiently via Monte Carlo tree search

Abstract

As a crucial step in machine learning, data preparation is the most time and energy consuming task for data scientists, entailing a number of data processing techniques to improve the performance of output results for ML models. However, end-to-end AutoML research focuses on automated machine learning pipelines consisting of algorithm selection and hyper parameter tuning, falling short in comprehensive automation of data preparation. In this paper, we propose Auto-DP, an MCTS-based framework for efficient and automated data preparation. To guide the search more effectively, a neural network is designed to estimate the subsequent maximum performance gain of each tree node. In order to reduce search space and improve system efficiency, two optimization strategies, meta-learning and accelerated training strategy, are used to determine the type and order of tasks in the data preparation process in advance, and speed up the pipeline creation process. We compare Auto-DP with the popular AutoML systems on 60 real datasets from OpenML repository. Auto-DP improves the Accuracy by up to 18.11 % on the classification task and reduces the Mse by up to 25.75 % on the regression task. Furthermore, it achieves a performance in 10 s that is better than what four popular AutoML systems achieve in 1 h.