World of Warcraft (WoW) is one of the most popular massively multiplayer online role-playing games (MMORPGs) having more than 10 million subscribers over the world. In order to engage and retain users understanding and predicting their behavior can be very useful for game developers. An important component of WoW are so-called guilds, which are social communities whose members can act together efficiently to accomplish more difficult goals and also provide a social atmosphere in which the game might be more entertaining. In this paper, we build predictive models to forecast which of the players will leave their guild in the close future. Our best model uses fuzzy c-means clustering to capture groups of similar guilds, that serve as the basis of an ensemble model, which computes predictions for each cluster separately and combines individual predictions into one final prediction using the memberships of the fuzzy clusters. Empirical analysis on WoW game data shows that our methods convincingly outperform the only existing method in the literature. To ensure transparency and reproducibility we publish the source codes of the research and also provide a Docker image, which makes it possible for anyone who has Docker installed to reproduce all of our results with a single command.

Web-based services that have become prevalent in people’s everyday life generate huge amounts of data, which makes it hard for the users to search and discover interesting information. Therefore, tools for selecting and delivering personalized contents for users are crucial components of modern web applications. Social recommender systems suggest items to users assuming the knowledge of the users’ social network. This new approach can alleviate the common weaknesses of traditional recommender systems, which completely ignore the users’ personal relationships in the recommendation process. In this paper, a social network based fuzzy recommendation technique is presented, which propagates information through the users’ social network and predicts how users would probably like a certain product in the future. Experimental results on a public dataset show that the proposed method can significantly outperform popular and widely used recommendation system methods in terms of recommendation coverage while maintaining prediction accuracy and performs especially well for cold start users, that have only rated a few items or no item at all previously.