Towards an Interpretable Metric for DOTA 2 Players: An Unsupervised Learning Approach

DOTA 2 is a Multiplayer Online Battle Arena game, which was released in 2013 and currently has approximately one million players worldwide. In this work, we approach the problem of finding a new metric to characterize and rank DOTA 2 players. We collect an extensive dataset of professional players and we show that the standard metric used in the community, called KDA, is not able to properly rank good players with different sets of skills. Combining unsupervised learning and evolutionary computing, we find a small set of players' attributes that we use to create a new metric, called GDM (General DOTA 2 Metric). In our results, we show that GDM is simple, easy to interpret and that it overcomes many of the issues related to using KDA to rank DOTA 2 players.