Deep Neural Networks for Optimal Team Composition

TL;DR

This paper investigates how teammates influence player performance in online games, proposing a deep learning framework to recommend optimal teammates and predict skill transfer effects, with applications demonstrated on Dota 2 data.

Contribution

It introduces a novel deep learning-based recommendation system for teammate selection and a measure of influence capturing skill transfer in team-based online games.

Findings

Teammate influence on performance can be effectively modeled.

Deep neural autoencoders achieve state-of-the-art recommendation accuracy.

Skill transfer dynamics are predictable using deep neural networks.

Abstract

Cooperation is a fundamental social mechanism, whose effects on human performance have been investigated in several environments. Online games are modern-days natural settings in which cooperation strongly affects human behavior. Every day, millions of players connect and play together in team-based games: the patterns of cooperation can either foster or hinder individual skill learning and performance. This work has three goals: (i) identifying teammates' influence on players' performance in the short and long term, (ii) designing a computational framework to recommend teammates to improve players' performance, and (iii) setting to demonstrate that such improvements can be predicted via deep learning. We leverage a large dataset from Dota 2, a popular Multiplayer Online Battle Arena game. We generate a directed co-play network, whose links' weights depict the effect of teammates on players' performance. Specifically, we propose a measure of network influence that captures skill transfer from player to player over time. We then use such framing to design a recommendation system to suggest new teammates based on a modified deep neural autoencoder and we demonstrate its state-of-the-art recommendation performance. We finally provide insights into skill transfer effects: our experimental results demonstrate that such dynamics can be predicted using deep neural networks.