Two Essays on User Behaviors in Online Games

Abstract

This dissertation consists of two essays on user behavior in online games. In the first essay, I identified multi-botting cheaters and measured their impacts using basic information in database such as user ID, playtime and item purchase record. I addressed the data availability issue and proposed a method for companies with limited data and resources. I also avoided large-scale transaction processing or complex development, which are fairly common in existing cheating detection methods. With respect to identifying cheaters, we used algorithms named DTW (Dynamic Time Warping) and JWD (Jaro–Winkler distance). I also measured the effects of using hacking tool by employing DID (Difference in Differences). My analysis results show some counter-intuitive results. Overall, cheaters constitute a minute part of users in terms of numbers – only about 0.25%. However, they hold approximately 12% of revenue. Furthermore, the usage of hacking tools causes a 102% and 79% increase in playtime and purchase respectively right after users start to use hacking tools. According to additional analysis, it could be shown that the positive effects of hacking tools are not just short-term. My granger causality test also reveals that cheating users activity does not affect other users' purchases or playtime trend. In the second essay, I propose a methodology to deal with churn prediction that meets two major purposes in the mobile casual game context. First, reducing the cost of data preparation, which is growing its importance in the big-data environment. Second, coming up with an algorithm that shows favorable performance comparable to that of the state-of-the-art. As a result, we succeed in greatly lowering the cost of the data preparation process by employing the sequence structure of the log data as it is. In addition, our sequence classification model based on CNN-LSTM shows superior results compared to the models of previous studies.