Predicting Customer Churn in World of Warcraft

TL;DR

This paper applies survival analysis and machine learning to predict customer churn in World of Warcraft, achieving high accuracy and providing insights into player behavior and game addictiveness.

Contribution

It introduces a combined use of survival analysis and classification models for churn prediction in an online game context, with a focus on a specific dataset from 2008.

Findings

Players tend to have long durations before churn, indicating high game addictiveness.

The best model achieved a 96% ROC AUC in churn prediction within six months.

Survival analysis helps understand player engagement over time.

Abstract

World of Warcraft is a massively multiplayer online video game released on November 23, 2004, by Blizzard Entertainment. In contrast with traditional games only having a single upfront fee to play, WoW also has a monthly subscription to play the game. With customer subscriptions in mind, we can apply the use of churn prediction to not only predict whether a customer will unsubscribe from the service but explore the user's playing behavior to obtain more insight into user playing patterns. The churn problem is somewhat complex due to the nature of not having a one size fits all solution, as different services define churn in a variety of ways. In this paper, we explore a dataset that focuses on one year from January 1, 2008, until December 31, 2008, as it highlights the release of a major content update in the game. Machine learning is used in two aspects of this paper: Survival Analysis and Binary Classification. Firstly, we explore the dataset using the Kaplan Meier estimator to predict the duration until a customer churns, and lastly predict whether a person will churn in six months using traditional machine learning algorithms such as Logistic Regression, Support Vector Machine, KNN Classifier, and Random Forests. From the survival analysis results, WoW customers have a relatively long duration until churn, which solidifies the addictiveness of the game. Lastly, the binary classification performed in the best performing algorithm having a 96% ROC AUC score in predicting whether a customer will churn in six months.