Dota2 pre-game prediction

• Main Author: Wu, Hao
• Other Authors: Tang Xueyan
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/138655
• Institution: Nanyang Technological University

Dota2 is a very popular Multiplayer Online Battle Arena (MOBA) game. This project aimed to use only Dota2 pre-game data, which is essentially the hero draft picked by the two teams, to predict the game result and the game duration using Deep Learning. Dota2 match data were collected through a third-party Application Programming Interface (API) and 181,717 matches were collected from 18 Jan 2020 to 26 Feb 2020. Data analysis was then conducted to investigate the influence of hero draft on game result and duration and feature engineering was also carried out to extract new features. Three models were built and game result prediction achieved an accuracy of 60.1% while game duration prediction achieved an accuracy of 59.6%. Game result prediction accuracy could be further improved to 67% if only considering predicted win rate above 60%, which makes more sense.