Efficient Feature Engineering and Lightweight AI Models for Predicting Lifestyle Indicators Using Lifelog Data

The importance of research on the utilization of lifelog data based on wearable devices is on the rise. Consequently, exploring efficient data utilization methods is becoming increasingly crucial. This study explores ways to efficiently use lifelog data to predict seven lifestyle indicators related to sleep, emotions, and stress. We propose a method for generating derived variables and feature engineering to identify key features for each indicator, and for employing predictive imputation for missing key features to streamline the given data. Using the proposed method, we were able to achieve 99.46% of the performance with data that included only about 15 % of the columns used by the top-performing model. This demonstrates the feasibility of conducting energy-efficient research by ensuring performance with minimal information.