Learning to recommend game contents for real-time strategy gamers

It is not surprising that there are currently many game videos, increasingly prevalent on the web, to watch professional player´s matches. Therefore, we need a kind of recommendation system to select the most preferable contents. Although there have been many personalization techniques proposed, there are few works on the recommendation of personal game contents. In this paper, we attempt to propose to use machine learning based on game contents with user´s explicit preference (like or dislike). Especially, we incorporate game domain knowledge on the design of features for learning. As a test bed, we select a famous real-time strategy game, StarCraft, because there are many game replays played by professional players. To generate training samples, each participant is invited to review replays and express his/her preference. Using the data, machine learning algorithms build a set of models to predict preference on new replays. For five participants, we labeled two hundred StarCraft replays. The exploratory study on the selection of the features and classification algorithms show the importance of the careful selection of them. The experimental results show that the proposed recommendation system can predict the users´ preference with an accuracy of 79% when we select appropriate models and features.