Deep Neural Networks for Optimal Team Composition.

Area of Science:

• Computational Social Science
• Human-Computer Interaction
• Machine Learning

Background:

• Cooperation is crucial for human performance, with online games serving as natural environments to study its effects.
• Player interactions in team-based online games can significantly influence individual skill acquisition and overall performance.
• Understanding teammate dynamics is key to optimizing player experience and learning in digital collaborative environments.

Purpose of the Study:

• To identify how teammates influence player performance in both short-term and long-term online gaming.
• To develop a computational framework for recommending teammates to enhance player performance.
• To demonstrate the predictability of performance improvements through deep learning models.

Main Methods:

• Utilized a large dataset from Dota 2, a popular Multiplayer Online Battle Arena (MOBA) game.
• Constructed a directed co-play network to quantify teammate influence on player performance.
• Developed a recommendation system using a modified deep neural autoencoder for teammate suggestions.

Main Results:

• Introduced a network influence measure capturing player-to-player skill transfer over time.
• The recommendation system achieved state-of-the-art performance in suggesting beneficial teammates.
• Experimental results confirmed that skill transfer dynamics in gaming are predictable with deep neural networks.

Conclusions:

• Teammate selection is a critical factor for optimizing performance and skill development in online games.
• The proposed computational framework and deep learning approach effectively predict and enhance player performance through strategic teammate recommendations.
• This research offers valuable insights into the mechanisms of skill transfer within collaborative online environments.