[4Xin1] Poster session 2

[4Xin1-75] Challenges of AutoML-based active learning towards data preparation with no expertise

〇Takumi Akazaki¹ (1.Fujitsu)

Keywords:AutoML, active learning, model selection

With growing demand for business application of machine learning, AutoML is actively studied as a solution for the shortage of data scientists. However, data preparation---the process of making AutoML input---still requires specialized skills. One of the techniques that assist the data preparation is active learning which performs the labeling process efficiently, but it also requires skill to decide suitable machine learning models for given data.
In this study, we explore the possibility of improving the efficiency of the labeling process without personal dependence by active learning based on AutoML, which dynamically determines the model. First, for each baseline model commonly used in AutoML, we empirically clarify its characteristics from the view point of the existing active learning framework.Furthermore, as a combination method of active learning and AutoML, we explore the suitable configurations of the acquisition function of active learning and the train-validation split in AutoML.