Player classification has recently become a key aspect of game design in areas such as adaptive game systems, player behaviour prediction, player tutoring and non-player character design. Past research has focused on the design of hierarchical, preference-based and probabilistic models aimed at modelling players' behaviour. We propose a meta-classification approach that breaks the clustering of gameplay mixed data into three levels of analysis. The first level uses dimensionality reduction and partitional clustering of aggregate game data in an action/skill-based classification. The second level applies similarity-based clustering of action sequences to group players according to their preferences. For this we propose a new approach which uses Rubner's Earth Mover's Distance (EMD) as a similarity metric to compare histograms of players' game world explorations. The third level applies a combination of social network analysis metrics, such as shortest path length, to social data to find clusters in the players' social network. We test our approach in a gameplay dataset from a freely available first-person social hunting game.