Mining player behaviour from a subscription‐based online game

Abstract

With an overabundance of data generated daily, data mining has become essential to gain the upper hand over competitors for every business. This alludes to the problem faced by this research, being how can a subscription‐based game, specifically World of Warcraft, leverage the available data to gain insightful information about their players’ behaviour to provide means to increase customer satisfaction and hence retain them from departing to numerous other games available. The solution presented in this research aims to predict the player’s development and provide an in‐depth analysis of different player groups to further the knowledge of the player’s behaviour. This is accomplished through three objectives; predicting both player activity within the game and player departure, along with identifying groups of players with similar interests. These are beneficial tools to provide the player with a tailor‐made experience and provide means for the company to know when to address a player before departing, providing marketing efforts to retain them. Research into churn prediction involved examining the performance of different binary classification models, namely Logistic Regression, Support Vector Machines, Decision Trees, Random Forest, and a Multi‐Layer Perceptron, while also determining the optimal look‐back to horizon ratio. The Random Forest model emerged as the best per‐ former, achieving an accuracy of 81% and an f1‐score of 83% with a horizon of 1 month and an accuracy of 79% and an f1‐score of 75% when predicting 3 months into the future by considering the player’s data from the past month. The investigation of the player’s development included comparing the performance of two ensemble models, AdaBoost and XGBoost, along with a Multi‐Layer Perceptron, to a naive baseline algorithm in predicting social and hostile interactions, map exploration, level, and playtime. The AdaBoost model was found to be the most effective, and all the machine‐learning approaches outperformed the baseline. In the player clustering task, various machine‐learning clustering techniques, including K‐Means, DBSCAN, and Gaussian Mixture Modelling, were employed to group players into distinct clusters. Separate prediction models were developed for each cluster, and the performance impact of less generalisation of each cluster was explored. Furthermore, an extensive analysis of each cluster was also conducted. The Gaussian Mixture Modelling technique yielded well‐defined clusters, representing beginners or traders, casual players, and hardcore players. Additionally, it was concluded that the separate prediction models did not aid in improving performance.