PENERAPAN METODE NAIVE BAYES UNTUK MEMPREDIKSI CUSTOMER CHURN BERDASARKAN KINERJA APLIKASI DUOLINGO DI INDONESIA

Abstract

The users of online learning applications such as Duolingo in Indonesia continue to increase, but retention challenges still arise due to customer churn. This research builds a churn prediction model based on application performance perception surveys with the Bernoulli Naive Bayes algorithm in the CRISP-DM framework. Data is obtained from 200 respondents through an online survey, then processed through data cleaning, removal of potentially leaky attributes, feature selection, encoding, to the formation of target variables. The model was evaluated with 5-fold cross-validation using accuracy, precision, recall, and f1-score metrics. The results showed an average performance of accuracy 0.8600, precision 0.8788, recall 0.8721, and f1-score 0.8717. Sensitivity analysis confirmed the model's performance remained stable even if one feature was removed, as the important information was spread across many variables. These findings can form the basis of a more comprehensive retention strategy, focusing on aspects such as language availability, clarity of learning progress, icon readability, streak status, and ease of menu navigation.

Keywords: Churn, CRISP-DM, Duolingo, Naive Bayes, Prediction