Utilizing Players’ Playtime Records for Churn Prediction Mining Playtime Regularity
Abstract : Churn prediction is an important topic in the freeonline game industry. Reducing the churn rate of a game signifi-cantly helps with the success of the game. Churn prediction helpsa game operator identify possible churning players and keep them engaged in the game via appropriate operational strategies, marketing strategies, and/or incentives. Playtime related featuresare some of the widely used universal features for most churnprediction models. In this paper, we consider developing newuniversal features for churn predictions for long-term playersbased on playtime. In particular, we measure playtime regularityusing the notion ofentropyandcross-entropyfrom informationtheory. After computing playtime regularity of players fromthe data sets of six free online games of different types, weleverage information from the playtime regularity in the form ofuniversal features for churn prediction. Experiments show thatthe proposed features are better at predicting churners comparedto the baseline features, implying that the proposed features couldutilize the information extracted from playtime more effectivelythan the related baseline playtime features.
• We model the long-term player in-game time spendingregularity based on the data of players’ in-game timespending distribution.
• We inspect churners’ and non-churners’ evolvement ofin-game time spending regularity across two free onlinegames.
• Then we the propose features based on playertime spending regularity of long-term players.
• We conduct experiments to evaluate our developed features across the two free online games’ data sets and show that these features could help achieve a better prediction performance than the baseline features.
• The combination of the static feature and rate feature ofentropy of thedailyplaytime distributions.
• We call thistype of features the1st type of the proposed features.
• The combination of the static feature and rate feature of cross-entropy of the daily playtime distributions.
• We call this type of features the 2nd type of the proposed features.
• The combination of the static feature and rate feature of entropy of the hourly playtime distributions.
• We call thistype of features the3rd type of the proposed features.
• The combination of the static feature and rate feature of cross-entropy of the hourly playtime distributions.
• We call this type of features the4th type of the proposed features.
• we propose prediction models using new features extracted from mining players’ in-game time spending regularity.
• After experiments are conducted, the corresponding result shows that our developed features are better at predicting churners, compared to the baseline features.
• Even without other game-specific features (e.g., login records, in-game events frequency, and payment information),we can still leverage the information from our examination of player in-game time spending regularity in the form of features and achieve satisfactory prediction performance.
? We evaluate the performance of the classifiers that aretrained with the data of the entiremplaying periods on thetest data set
? Wetrain classifiers using the training data set and evaluate theirperformance using the test data set.
? LSTM is trained usingmean squared error as a loss-function and an Adam optimizer.
|