Skripsi
ANALISIS PSIKOLOGI PENGGUNA TERHADAP FITUR GAMIFIKASI DUOLINGO MENGGUNAKAN MACHINE LEARNING
This research analyzed the psychological impact of Duolingo's gamification features through user reviews on the Google Play Store. A total of 5,000 reviews were collected and extracted into eight gamification features using a rule-based keyword extraction approach, including penalty, streak, xp_level, reminder_notification, leaderboard, reward_gem, daily_challenge, and badge_achievement. Psychological labels were determined based on star ratings, categorized into Motivated (rating 4–5), Frustrated (1–2), and Neutral (3). Extreme class imbalance was addressed using SMOTE for Naïve Bayes and class weights for Random Forest, SVM, and Logistic Regression. Results showed that the penalty feature was the most dominant (141 reviews) and became the most influential feature in the feature importance analysis with a score of 0.513. A performance comparison of the four algorithms resulted in nearly identical accuracy (92.6%–92.8%). Naïve Bayes achieved the highest accuracy but failed to detect the Neutral class, while SVM and Logistic Regression with class weights successfully detected 1 out of 12 Neutral reviews (0.08% recall). The Frustrated class posed the greatest challenge with only 19% recall across all models. These findings confirm that penalty is the most crucial feature, and class weight is more effective than SMOTE in handling extreme data imbalance. Practical implications for Duolingo developers include prioritizing improvements for penalty and daily challenge features, and adopting class weight as the primary strategy for handling imbalance.