Temporal Churn Prediction in Mobile Game with Fuzzy Behavioral Augmentation and SHAP Explainability

The entertainment industry covers numerous fields of business, such as telecommunication, television, music, video games, and event management. Out of these markets, the gaming industry has been growing significantly due to its high revenues, a substantial number of players, and overall popularity among customers worldwide. Player churn prediction may inform retention decisions, although predictive performance alone does not establish that an intervention improves retention. This study uses fuzzy behavioral modeling and SHAP-based explanations to examine temporal churn prediction in a single mobile word puzzle game. We examine whether fuzzy representations of player activity improve the prediction of future inactivity in a mobile word puzzle game. The dataset contains 2602 users, 46,132 session identifiers, and 50,014 session-level observations collected over 86.41 days. At each prediction date, eligible players have been active during the prior 14 days; churn is defined as no recorded activity in the following 14 days. Historical training observations accumulate across dates, and all preprocessing is fitted on training data. Six ML classifiers are evaluated using 183 base features, 199 fuzzy-augmented features, and filtered versions of both sets. Filtering retains 85 base or 91 augmented features for the original test set. Two historical validation transitions are followed by test periods containing 144 and 101 players, respectively, with 245 observations from 204 distinct players in the combined assessment. The Logistic Regression configuration selected by historical validation achieves original-test balanced accuracy of 0.608, AUC of 0.700, and an F1-score of 0.643. Its combined balanced accuracy is 0.620 (95% player-cluster bootstrap interval 0.543–0.692). Paired ablations show that fuzzy augmentation does not provide a consistent advantage across classifiers and evaluation settings; the effects of feature reduction also vary. These performance estimates are specific to this dataset, its recently active players, and the stated temporal protocol. They do not establish general predictive performance or the usefulness of fuzzy augmentation in other games, populations, or observation periods.