Representing Player Behaviour via Graph Embedding Techniques: A Case Study in Dota 2

We explore the use of graph embedding techniques to represent the player behaviour that is expressed in the logs of video games. While such logs hold data that could be useful for personalization, the data is often poorly structured for use with Artificial Intelligence systems and its dimensionality is often high. By using a graph to structure the logs and applying embedding techniques to reduce their dimensionality, a compact vector representation can be obtained that preserves some of their semantics. To explore the potential value of this approach, we obtained gameplay logs from over 3000 matches of Defense of the Ancients 2 (Dota 2) and compared 13 parameter variations of three different embedding techniques: NODE2VEC, LINE, and TGN. Our analysis considers the effects of embedded vector size, dataset size, a step size used for updating vectors as a game proceeds, and different types of player interaction. The results show that NODE2VEC outperforms the other techniques on 7 of the 13 variations that we tested, and that removing one type of player interaction can make it easier to predict the others.