Adaptivity of Card Recommendation Systems for Legends of Code and Magic

The complexity of online games asks for support of new players, e.g., by recommending deck improvements in case of Collectible Card Games. But many online games regularly change the properties of existing game elements to either balance the game dynamics or to keep it interesting for its community. To successfully offer such a recommendation service, it is inevitable that they adopt to such changes. We study four card recommendation systems, that provide recommendations based on logistic regression models trained from game log data. The recommendation systems are compared based on their models’ ability to adapt to different scenarios of change: New cards are introduced, existing cards are changed, and opponent behavior is changed. As expected, when updating the recommendation systems without training on new data, recommendation accuracy was impaired, with the severity depending on model features and the kind of change. Using a genetic algorithm, a simulation of a player community was implemented and recommendation system accuracy and adaptivity were explored in a competitive environment. In the experiments, systems with high accuracy have proven more robust against change even if it took them more data (and thus time) to adapt.