Good Game Well Played: Predicting DotA 2 Match Outcomes Using Machine Learning Classifier Algorithms

Authors: Joshua Cuballes, Rosely Pena, Martin Salvano, Jonathan Uy

Executive Summary

In the industry of e-Sports, the increased popularity of Multi-Online Battle Arena video games such as DotA 2 led to the emergence of Big Data. Big data are extremely large data sets that can be analyzed to reveal patterns, trends and associations, particularly with respect to human behavior and interactions. DotA 2 is a video game developed and published by Valve, a multiplayer online battle arena consisting of two teams, Radiant and Dire. This game attracted a lot of audience hence increasing the stakes which includes e-Sports betting. Given this situation, this study aims to predict the outcome of the game even before the game starts. Using data from OpenDota, an open-source blog for DotA 2 data, predictive models were built using heroes as the main features. The models used to predict were Decision Tree, Random Forest and Gradient Boosting Machine classifiers. The baseline model, DT, shows a 52% accuracy but was increased to 64% using the GBM classifier with hyperparameter tuning. Results show that the top three heroes highly influencing the game were Spectre, Omniknight, Zeus. Frequent Itemset was also used for further validate the results of the ML models and results show that the frequently used heroes of the winning team are not the heroes that drive the outcome of a winning game. Findings of the study can be used by different audiences such as researchers who are starting to explore the potential of the data as well as e-Sports enthusiasts especially those who spend money on e-Sports betting. For further studies, it is recommended to use Deep Learning models to improve the accuracy and to add features such as banned heroes and items purchased during the game.

Note: If you wish to have a copy of the technical paper (PDF), data and the code used in this project, you may send me an email or you can also drop your message via LinkedIn.