GAIM: Game Action Information Mining Framework for Multiplayer Online Card Games (Rummy as Case Study).

We introduce GAIM, a deep-learning analytical framework that enables benchmarking and profiling of players, from the perspective of how the players react to the game state and evolution of games. In particular, we focus on multi-player, skill-based card games, and use Rummy as a case study. GAIM framework provides a novel and extensible encapsulation of the game state as an image, and uses Convolutional Neural Networks (CNN) to learn these images to calibrate the goodness of the state, in such a way that the challenges arising from multiple players, chance factors and large state space, are all abstracted. We show that our model out-performs well-known image classification models, and also learns the nuances of the game without explicitly training with game-specific features, resulting in a true state model, wherein most of the misclassifications can be attributed to user mistakes or genuinely confusing hands. We show that GAIM helps gather fine-grained insights about player behavior, skill, tendencies, and business implications, that were otherwise not possible, thereby enabling targeted services and personalized player journeys.