Using Machine Learning to Optimize Team-Based eSports Outcome Prediction

Abstract

In this thesis, we attempt to create a machine learning model that predicts the outcomes of professional MOBA matches. To possibly address limitations of past contributions to the field, we incorporate both prior and real-time features into our analysis, and also perform spectral clustering in an attempt to better understand what makes successful team combinations. For our dataset of 24,307 professional-level matches, our composite regression model reaches accuracy as high as 95.33% at minute 45, beginning from prior based accuracy of 68.70%. A clustering algorithm performed on a dataset of 4,337,598 matches produces 23 hero clusters for examination.