[4T2-GS-10] AI application

[4T2-GS-10-01] Dropout Prediction using Positive and Unlabeled data to improve student satisfaction at an online school.

〇Miki Katsuragi¹, Kenji Tanaka¹ (1. The university of Tokyo)

[[Online]]

Keywords:Dropout Prediction, Learning from Positive and Unlabeled data, Explainable AI

In this study, we analyzed data on the daily learning status of students enrolled in an online school and attempted to predict school withdrawal using machine learning. Specifically, we attempted to predict the withdrawal of students after one month based on the students' course data and Slack conversation data for the past three months. In other words, it is necessary to consider not only the timing of withdrawal prediction, but also the possibility of future withdrawals. In order to solve this problem, instead of simply using the binary classification of Positive and Negative, we took the approach of Learning from Positive and Unlabeled data, and made predictions based on the classification of Positive and Unlabeled. As a result, we were able to improve the predicted return rate of students who took a leave of absence or dropped out of school from 63% to 72%.