[2D5-GS-2] Machine learning: Statistical learning

[2D5-GS-2-02] Dynamic Feature Selection Using Features of Features

〇Takahashi Katsumi¹, Koh Takeuchi¹, Hisashi Kashima¹ (1. Kyoto University)

Keywords:Feature Selection, Deep Learning, Active Learning

In general, we input values of features into a machine learning model to obtain an output. In many real-world problems, there will be a cost to measure a value of features. Therefore, in order to reduce the measurement cost, research has been conducted on methods that dynamically select features to be measured. However, existing methods assume that the set of features is always same, so they cannot handle cases where the set of features that can be measured changes from instance to instance. In this study, we propose a new method that assumes prior information about conventional features as "features of features" and dynamically selects features based on this prior information and the set of measured features. Experiments on several datasets confirm that the proposed method appropriately selects features based on the prior information even when the set of features that can be measured changes from instance to instance.