An Empirical Study on Team Formation in Online Games

Essa Alhazmi, Sameera Horawalavithana, Adriana Iamnitchi, John Skvoretz, Jeremy Blackburn

Abstract

Online games provide a rich recording of interactions that can contribute to our understanding of human behavior. One potential lesson is to understand what motivates people to choose their teammates and how their choices lead to performance. We examine several hypotheses about team formation using a large, longitudinal dataset from a team-based online gaming environment. Specifically, we test how positive familiarity, homophily, and competence determine team formation in Battlefield 4, a popular team-based game in which players choose one of two competing teams to play on. Our dataset covers over two months of in-game interactions between over 380,000 players. We show that familiarity is an important factor in team formation, while homophily is not. Competence affects team formation in more nuanced ways: players with similarly high competence team-up repeatedly, but large variations in competence discourage repeated interactions.

Introduction

Understanding how humans self-organize into teams is a fundamental challenge in sociology and organizational behavior. While real-world team formation is often difficult to track at scale, online video games offer massive, meticulously recorded environments of goal-oriented human interaction. Unlike many modern competitive games that use algorithmic matchmaking, Battlefield 4 (BF4) utilizes a "player choice" system. In BF4, players manually select servers and can form micro-teams called "squads" within larger 32-player teams. This environment provides a unique opportunity to quantify how factors like familiarity, homophily, and competence influence the active choice to collaborate.

Data and Methodology

The study utilized a custom longitudinal dataset crawled from the GameME API over an 81-day period (May to July 2016). The final cleaned dataset includes 384,066 distinct players participating in 60,410 matches across 63 servers in 7 countries.

Distribution of player activity showing the number of players versus the total number of games played.

The researchers mapped in-game data to three core sociological concepts:

  1. Familiarity: Measured by the frequency of two players appearing in the same match on the same or opposing teams.
  2. Homophily (Similarity): Measured by shared country of origin or shared "clan" tags.
  3. Competence: Measured by in-game server rank, kill/death ratios, and skill progression.

Squad formation was treated as a binary classification problem. Given the sparsity of the data (most player pairs never squad up), the Synthetic Minority Over-sampling Technique (SMOTE) was used to balance the training data for machine learning models, including Random Forest and Logistic Regression.

Key Insights: Drivers of Squad Formation

The analysis revealed that the unit of cooperation—the squad—is typically small, with most players grouping in pairs or small clusters.

Cumulative distribution function (CDF) of squad sizes, showing that the vast majority of squads consist of 5 or fewer players.

Familiarity and Competence

The strongest predictor of team formation is familiarity. The more games two players play together, the exponentially more likely they are to actively choose to join the same squad. Using only familiarity features, a Random Forest model predicted squad co-membership with over 80% accuracy.

Regarding competence, the data suggests players prefer peers over "carries." Players naturally form squads with others of similar skill levels. The median standard deviation of player ranks within real squads was approximately 1,800, compared to over 4,000 in randomly generated squads. This suggests players avoid the potential toxicity of high-skill gaps.

The Myth of Homophily

Surprisingly, homophily played a minor role at the squad level. While servers are often geographically clustered due to network latency, players do not actively segregate into squads based on country of origin. In fact, two-thirds of all observed squads contained players from at least two different countries.

Conclusion

This empirical study demonstrates that familiarity and competence-matching are the primary drivers of team formation in player-choice environments. These behaviors appear universal across different server populations. Future work could refine the definition of familiarity by accounting for the quality of past interactions, such as shared victories or defeats, to better understand the dynamics of long-term cooperative bonds.

Summary updated

Abstract

Online games provide a rich recording of interactions that can contribute to our understanding of human behavior. One potential lesson is to understand what motivates people to choose their teammates and how their choices leadto performance. We examine several hypotheses about team formation using a large, longitudinal dataset from a team-based online gaming environment. Specifically, we test how positive familiarity, homophily, and competence determine team formationin Battlefield 4, a popular team-based game in which players choose one of two competing teams to play on. Our dataset covers over two months of in-game interactions between over 380,000 players. We show that familiarity is an important factorin team formation, while homophily is not. Competence affects team formation in more nuanced ways: players with similarly high competence team-up repeatedly, but large variations in competence discourage repeated interactions.
BibTeX
@inproceedings{alhazmi2017empirical,
  author = {Essa Alhazmi and Sameera Horawalavithana and Adriana Iamnitchi and John Skvoretz and Jeremy Blackburn},
  title = {{An Empirical Study on Team Formation in Online Games}},
  booktitle = {ASONAM},
  year = {2017},
  url = {https://openreview.net/forum?id=Yk6drxJYVH},
}
Figure