Fig 1.
Time diagram of a player’s play log data.
The player has played n = 8 times. The first three plays occurred within two days of the first play. The next two plays occurred in the following two days.
Fig 2.
Two churners and two active users.
Each row describes play log records of a user, and a user with no play log record in the churn prediction period is defined as a churner. When desired, a larger CP can be chosen for defining churn.
Table 1.
Three games and their data.
Table 2.
Principal attributes of Game #1.
Table 3.
Principal attributes of Game #2.
Table 4.
Principal attributes of Game #3.
Table 5.
Common features of the three games.
Table 6.
Game-specific features of Game #2 and Game #3.
Fig 3.
Evaluation of single features.
(a) Based on correlation with churn. (b) Based on gradient boosting performance.
Table 7.
Overall average rankings for single features.
Table 8.
The underlined features are the top 2 from the single-feature ranking.
Fig 4.
AUC vs. number of utilized features.
Increasing number of features has only a marginal effect on AUC performance, especially after 2~3 features are included.
Fig 5.
Input to the deep learning algorithms.
For each user, the value in each time frame is the sum of all scores that were recorded during the time frame.
Table 9.
Performance (AUC) comparison of three conventional algorithms and two deep learning algorithms (OP = 5 days, CP = 10 days).
Fig 6.
AUC performance as a function of OP and CP selections. The churn prediction performance is improved by increasing number of observation days and by decreasing number of churn prediction days.
(a)(c)(e) AUC vs. (OP,CP) for Game #1, #2, and #3, respectively (b)(d)(f) AUC vs. CP for four fixed OP values for Game #1, #2, and #3, respectively.