Comparing the Structures and Characteristics of
Different Game Social Networks - The Steam Case
Enrica Loria
ISDS Institute
Graz University of Technology
Graz, Austria
Alessia Antelmi
Dipartimento di Informatica
Universit
`
a degli Studi di Salerno
Salerno, Italy
Johanna Pirker
ISDS Institute
Graz University of Technology
Graz, Austria
Abstract—In most games, social connections are an essential
part of the gaming experience. Players connect in communities
inside or around games and form friendships, which can be
translated into other games or even in the real world. Recent
research has investigated social phenomena within the player
social network of several multiplayer games, yet we still know
very little about how these networks are shaped and formed.
Specifically, we are unaware of how the game type and its
mechanics are related to its community structure and how those
structures vary in different games. This paper presents an initial
analysis of Steam users and how friendships on Steam are
formed around 200 games. We examine the friendship graphs
of these 200 games by dividing them into clusters to compare
their network properties and their specific characteristics (e.g.,
genre, game elements, and mechanics). We found how the
Steam user-defined tags better characterized the clusters than
the game genre, suggesting that how players perceive and use
the game also reflects how they connect in the community.
Moreover, team-based games are associated with more cohesive
and clustered networks than games with a stronger single-player
focus, supporting the idea that playing together in teams more
likely produces social capital (i.e., Steam friendships).
Index Terms—Social Network Analysis, Steam, Games, Player
Network
I. INTRODUCTION
Well-being [1], increased engagement [2], [3], longer-term
retention [4], and a sense of connectedness [5] are established
benefits deriving from playing together. Researchers increas-
ingly show that playing video games is a social activity [1],
[5], [6]. Players gather in online and offline communities
revolving around games, not only relying on the embedded
social mechanics of the game but also on social media [7]
and other forums [8]. This enhances the social aspect of single-
player games, for which a community of affectionate players
can also exist and connect. Studying the player community
provides much information on how the game is perceived and
also on the gameplay experience. Consequently, the analysis
of the player social network (either implicit and explicit) is
a practice that is catching on among researchers and game
analysts. Most of the research on Social Network Analysis
(SNA) in games is in the realm of strongly multiplayer
games (often team-based), such as League of Legends [9],
World of Warcraft [5], and Destiny [3], [10], hence neglecting
single-player games communities. Those works are generally
interested in connecting social interaction patterns to players’
activity and engagement in the specific context without analyz-
ing the community at a higher level and making a cross-game
comparison. Players connect in the game, social platforms, and
game providers. As a result, the investigation of the social
nature of games is more complex than just analyzing in-
game social networks. Nevertheless, we are still unaware of
how and whether the characteristics of games reflect on the
way players connect and which are the design elements or
mechanics fostering more cohesive communities. Researchers
argue that games build friendships, which have value in the
real world and are transferred across games. However, a still
unanswered question is: how does the type of game impact
the formation of those friendships?
We analyzed a game provider, Steam
1
as a step towards
the analysis of player friendships. The Steam platform is
a great incubator of social relationships among players due
to its hybrid nature combining functionality and few social
media elements. Steam is primarily a game library where
users can buy and store their games, build a profile, and share
their activities and achievements. Additionally, Steam allows
players to institute explicit friendship links with each other
and obtain updates on the actions of friends (e.g., playing,
reviewing, and buying new games) and progress. Specifically,
we collected and analyzed parts of the Steam friends network
and filtered it by each game played from the population
sample. We then clustered the game friends sub-graph in
six clusters. We characterized the clusters according to their
genre and user-defined tags and the network properties (e.g.,
clustering coefficient, degree distribution, and centralization).
Research Goal and Contributions
In summary, the Games User Research (GUR) literature
provides evidence on the importance of sociality in games and
the community built around games. Yet, the existing investiga-
tions into the social network of the players are conducted on
specific games and with clear objectives in mind (e.g., studying
retention, social influence, and engagement). Consequently,
we are unaware of the social network characteristics that can
1
https://steampowered.com
978-1-6654-3886-5/21/$31.00 ©2021 IEEE
arXiv:2106.12264v2 [cs.SI] 24 Jun 2021
be used to describe different game types. This study aims
to perform an exploratory investigation to identify how the
game characteristics, expressed by Steam genres and user-
defined tags, relate to the shape player community. Our results
connect some tags to specific network properties, contributing
to our understanding of how users conceive the games (i.e.,
tag them) reflects on the specific game community structure.
For instance, single-player games have scattered networks,
whereas multiplayer games show more connected graphs,
especially when team-based. Those networks are often scale-
free and thus resemble the degree distribution of social media
networks. Unlike tags, the game genres cannot be related
to community characteristics. As a result, our contribution
to the GUR community is twofold. First, we show how the
definition of games through tags by the players (in the Steam
platform) can be more meaningful for the understanding of
the community than the genre originally associated with the
game by Steam (or the designers). In other words, the way
the game is perceived and used by the players provides a
good picture of the actual player network shape. On the other
hand, the game genre does not provide additional information.
Second, we show how including a team-based (and often local)
multiplayer mechanics in games would more likely produce a
cohesive, connected community of friends that resembles (in
structure) social media networks. This finding further supports
the conception of multiplayer (team-based) games as social
incubators.
II. RELATED WORK
People seek relatedness in any kind of activity [11], and
very frequently do so in games [12]. Games embed a degree
of social connection by their nature, and do so either directly or
indirectly, in a way different to that in non-gameful actions.
Players are more likely to be kept engaged and to keep on
playing when their need to belong is fulfilled [2], as the
nature of in-game relationships can impact their behaviors
and participation [13]. Social play can produce feelings of
well-being and bring about a performance increases in the
players [3]. Well-designed social game mechanics can result
in the players having a strong motivation to complete their
tasks and to be retained for longer in the game: they are more
motivated to have success [6]. Within virtual environments,
players can form long-lasting friendships within games, which
can continue not only in the real world but also in other
games [14].
In this section, we review the literature on SNA in games
to show how researchers examined player networks within
specific games. We then focus on the studies analyzing the
game provider Steam, and identify a research opportunity in
the investigation of the Steam friends network filtered by each
of the games studied.
Social Network Analysis in Games
Researchers and designers can rely on SNA to monitor
the status of in-game social interactions. Player social rela-
tionships can be modeled using graphs, which successfully
represent interaction patterns among a group of people [15].
Social networks manifest when directed (e.g., following an
account) or indirected (e.g., befrending another user) social
interactions are allowed. Social media and standard Online So-
cial Networks (OSNs) explicitly define connections among the
users, linked because they are related or share interests. On the
other hand, despite being originated from indirect connections,
implicit social networks are a rich source of information, and
they may thus enforce similar social rules. For instance, online
multiplayer games, which are a social phenomenon, encourage
social interactions. These interactions, or relationships, can
also be interpreted as social networks, thus being modeled
using traditional SNA techniques. Online multiplayer games
convey information on the social aspect of gaming [16] and
help understanding social relationships in a highly digitalized
world [17]. The study of how players socialize through games
can lead to better social environments in games [17], [18].
Studying the player network in the form of a graph for
example, can highlight how players’ activity is reflected in the
experiences of others [19] and how the permanence of certain
players can condition others’ behaviors [19]. An inspection of
the player network hinted that spending more time in teams
is not a synonym of being more social, since the interests
of players in social interactions may be purely functional to
the game [18], [20]. Similarly, toxic interaction patterns may
emerge from the analysis [21]. Although the employment of
SNA in games is still young, researchers have already analyzed
the social roles of players. Group formation represents not
only a pillar of the player community [17] but sometimes
loyalty to the guild also led players to prioritize its growth
over their own personal interests [22]. The team organization
and connectedness also benefit individual performance [9]
and retention [3]. Multiplayer, or social, games foster social
relationships by their nature and can thus also be modeled in a
graph. Researchers studied groups and community [23], inves-
tigating the impact of social structures in gameplay [3], [24].
Properties of groups and guilds, for example, are indicators of
players’ in-game activity and retention [4], [25]. Although the
player communities formed around games comprise important
information about the social dynamics that occur [8], they say
much concerning the network composition of the game under
investigation.
Data Analysis on Steam
Games are multifaceted and as such they engage and
connect users in different ways. Consequently, analyzing inter-
action patterns within games provides a context-specific view
of players’ social dynamics. Conversely, investigating a game
provider could give us a higher-level perspective of players’
relationships.
Steam is an online distribution platform for video games.
Games on Steam are usually uploaded by the game developers,
game studios, or publishers. Each game is presented on a
store page. This page contains information about the game
such as screenshots, descriptions, developer and publisher
information, release year, genre, tags, a list of user reviews, a
user rating obtained from the user reviews. Steam also enables
connecting to other players. We can gain various insights about
players and games by looking at this heterogeneous dataset.
Several research teams have used Steam data to understand
player preferences [26] and play patterns [27]. For instance,
clustering playtime allowed the identification of connections
among game genres and how players span across different
genres [27]. The social structure of the Steam player network
has also been analyzed over the years. The Steam friends
network was a modest loosely connected graph in 2011 [28]
and grew substantially in 2016 [29], in the course of which
players were found to befriend other players who they found
similar to themselves and who also favored social games.
This indicates that players are indeed interested in social
interactions in games and game platforms. In the context of
Steam, researchers mostly studied the activities and behaviors
of players to achieve a greater knowledge of the individuals.
However, little is known on how players engaging in the same
game are connected and the relationship among those sub-
communities, and the specific properties of the games.
III. DATA
In this section, we detail the construction of the data set used
to perform our study. Starting by describing how we chose the
Steam users and how we built their friendship network, we
then detail how we picked the games we considered in this
analysis and their induced friends’ network.
A. The Steam Friendship Graph
We modeled the Steam friends network as an undirected,
non-weighted graph, in which the nodes are the players and
the edges represent a friendship status between the two nodes.
In the following, we detail the data set construction process.
Our intention was to increase the likelihood of the presence
of active users. The rationale behind this choice is that regular
players actually build a community around a given game or a
set of games. On the other hand, the inclusion of inactive play-
ers in the friendship network would only modify its structure
without contributing to the underlying social dynamics.
• Step 1. We randomly collected a seed set of 1k users
from the authors of the reviews of the top 100 “New and
Trending” games on the 10th of April 2020. We collected
the seed users’ friends and built a first friendship network,
resulting in a sample of about 50k nodes.
• Step 2. We then extracted the largest connected compo-
nent (LCC) of the first friendship network, consisting of
2.8k users.
• Step 3. We further retrieved the friends of the 2.8k users
in the LCC, obtaining a second friendship network with
240k nodes.
• Step 4. For the last iteration, we also retrieved the friends
of the newly added nodes. However, we did not store any
new node (i.e., player) in the graph; we only integrated
the network with the missing edges. Finally, we removed
the users with private profiles.
TABLE I: Distribution of the size of the friendship graphs.
Min 25% 50% 75% Max Mean Std
#nodes 270 369 505 833 17, 273 879.79 1, 445.36
#edges 7 38 91 217 32, 507 478.60 2, 413.29
The network obtained by this means counted 191, 479 nodes
and 1, 242, 093 edges. We then collected daily updated infor-
mation about the activities of each player(node) in the network
and did this in the form of time spent playing each game they
owned. The observation period covered five weeks, from April
13th 2020 to May 17th 2020. Given that we consider users to
be active if they have played at least one game during the
period of observation, we found that only 51k out of 191k
players were active during the five weeks we crawled their
activity. We removed all inactive users from the friendship
graph, along with the nodes that became disconnected. This
resulted in a final friendship network of 39,354 users and
218,432 edges.
B. Games’ Friendship Graphs
We analyzed the friendship network induced by each Steam
game to study whether game characteristics, such as genres
and tags, reflect on friendship ties of its players. We picked
the 200 most frequently played games within our dataset
and for each of these games we consider a subgraph of
the friendship network comprised only the users who have
played that game. Thus, we have an edge between two players
in each game subgraph if they have both played the given
game. We considered only the top 200 most-played games to
ensure a game-induced network of at least 250 nodes. Table I
summarizes the size distribution of the 200 game networks in
terms of the number of nodes (players) and edges (friendship
ties).
IV. METHOD
This section describes the method we followed to detect
any pattern in the friendship network induced by specific game
characteristics, detailing how we grouped similar networks and
the metrics to characterize each group.
A. Graph Clustering
Our analysis’s first step was to group all networks with
similar structural characteristics to evaluate whether different
games with analogous peculiarities generate similar linking
patterns in their induced friendship network. Several clustering
algorithms have been proposed in the literature to accomplish
this task (i.e., grouping items). However, they all deal with
data described by numerical vectors.
One way to describe graph-structured data through a numer-
ical representation is by engineering handcrafted features. The
key characteristics of the networks are thus manually chosen
and their values will compose the corresponding numerical
vector. Although it is straightforward, this approach may
introduce a bias towards the selected features and propagate it
in the results. Further, latent patterns encoded in the network
may be overlooked or missed. For this reason, we opted for a
graph embedding technique to transform a friendship network
in its numeric representation [30]. Given a graph G = (V, E)
and a predefined embedding dimension d, with d |V |, the
problem of graph embedding is to map G into a d-dimensional
space (a.k.a. latent space), in which the structural properties of
G are preserved as much as possible. Following this definition,
each graph is represented as either a d-dimensional vector (for
a whole graph) or a set of d-dimensional vectors. Each vector
represents the embedding of part of the graph, such as nodes,
edges, or substructures.
As we were interested in clustering graph structures, we
used graph2vec [31], a neural-network-based technique for
whole graph embedding. The underlying idea of this graph
representation learning process is to yield similar embeddings
for structurally similar graphs. Based on the document embed-
ding model [32], graph2vec sees each graph as a document
and the rooted subgraphs around every node in the graph as
words that compose the document. The graph2vec embeddings
are task agnostic and as such can be directly used across all
analytics tasks involving whole graphs.
Once we obtained the 200 vector embeddings - one for
each game friendship graph - we fed them to the K-means
clustering algorithm. We used the Silhouette analysis [33],
giving information about the separation distance between the
resulting clusters and the distortion index, counting the sum
of squared distances of samples to their closest cluster center,
to select the best number of clusters.
B. Clusters Characterization
As a means of characterizing the properties of each cluster,
we defined two sets of features considering i) the structural
characteristics of the game friendship networks within each
group and ii) the metadata associated with each game. This
approach is similar to the approach presented by Overgoor
et al. [34], where the authors analyze the impact of U.S.
college dynamics over students’ Facebook friendship relations.
From the structural point of view, we analyzed the following
features:
• size - number of nodes;
• edge density - the share of node pairs that are connected;
• mean and variance of node degree distribution;
• degree assortativity - which measures the similarity of
connections in the graph with respect to the node degree;
• group degree centralization - which equals 1 in a star
graph, 0 in a circle graph where all degrees are equal;
• group betweenness centralization - which reaches its
maximum value 1 for the star graph;
• average clustering coefficient - which measures the con-
nectedness of a network counting the number of triangles
a node is actually involved in overall possible triangles
in its neighborhood;
• number of connected components and size of the LCC;
• modularity - which measures how good a given graph
partition is based on the number of inter- and intra-
community edges;
Fig. 1: The Elbow Method showing the optimal k.
• percentage of the game networks following a power-law
distribution.
Regarding the game metadata, we considered:
• Steam-defined game genres;
• user-defined tags, indicating game mechanics, genres,
themes, or attributes. Generally, those tags can be any
term or phrase
2
. As we had an excessive number of tags
for each game (mean = 38.57, and std = 30.22), we
used Term Frequency Inverse Document Frequency [35]
(TD-IDF) to obtain a more representative list of user tags
for each game. TD-IDF is a numerical statistic intended
to reflect how important a word is to a document in
a collection or corpus. Hence, we used this metric to
retrieve the most characterizing tags (words) for each
game (document) in our game dataset (corpus).
V. RESULTS
A. Games’ Friendship Graphs Clustering
To obtain the network embeddings, we use the implemen-
tation of graph2vec at the following link
3
. We run the script
using the default parameters suggested by the authors to get
an 8-dimension graph embedding. Regarding the execution of
the clustering algorithm, we use the sklearn Python imple-
mentation of K-means
4
. We iterated the algorithm initializing
k from 2 to 10. Figure 1 shows the distortion index associated
to each value of k. The elbow does not clearly denote the best
k (either 6, 7, or 8 are valid choices). We decided to set k = 6,
since increasing k led to a higher number of clusters of size 1.
B. Characterizing the Game Clusters
Having clustered the games into six clusters (or groups), we
then analyzed the games’ properties to characterize each group
(see Section IV-B). The structural properties of each cluster
are summarized in Table II.
• Cluster 0. In this cluster we have 106 games, which
show the most dense networks, but with a low clustering
coefficient. The network is comprised of many connected
components and the largest of these covers about 10%
of the nodes. This division in many components is also
reflected by the modularity score, which is among the
2
https://store.steampowered.com/tag/
3
https://github.com/benedekrozemberczki/graph2vec
4
https://scikit-learn.org/stable/modules/generated/sklearn.cluster.KMeans.html
TABLE II: Network basic statistics, computed as the average value of each network graph in the cluster.
Cluster nodes density mean deg std deg avg clust #CC %LCC modularity assortativity %pl deg centr betw centr
#0 (106) 742 0,0014 60,95 145,10 0,030 583,92 0,129 0,727 0,110 63% 5,02E-05 0,016
#1 (1) 5.299 0.0005 17,78 147,21 0,067 2260 0,495 0,637 -0,06 100% 1,05E-05 0,171
#2 (72) 395 0,0005 84,58 132,22 0,009 363,04 0,021 0,863 0,169 88% 3,06E-05 0,001
#3 (13) 2.741 0,0004 74,49 271,93 0,042 1773,78 0,179 0,769 0,093 62% 6,76E-06 0,019
#4 (1) 5.419 0,0004 27,23 214,70 0,052 2964 0,325 0,481 0,017 0% 6,68E-06 0,068
#5 (1) 17.273 0,0002 48,65 390,76 0,107 3783 0,711 0,617 0,062 100% 1,17E-06 0,059
(a) Cluster 0 (106 games)
(b) Cluster 1 (Rocket League)
(c) Cluster 2 (72 games)
(d) Cluster 3 (13 games)
(e) Cluster 4 (Wallpaper Engine)
(f) Cluster 5 (CS:GO)
Fig. 2: Word clouds for the 5 game clusters, displaying the frequency of the user-defined tags in Steam.
highest. The games in the cluster present a centralized
structure, in terms of degree. The nodes, on average,
present a high degree distribution but very variable
(mean = 60, and std = 145).
• Cluster 1. This cluster includes one game: Rocket
League. The network is made of about 5k nodes, is not
particularly dense, and has the lowest degree distribution
(mean = 17 and std = 147). The degree distribution
follows the power-law distribution, and thus the network
is scale-free. Although many communities exist in this
network, the largest connected component accounts for
half of the nodes. Finally, the network is centralized, in
terms of betweenness centrality.
• Cluster 2. In this cluster, we have 72 games, which have
(on average) the smallest network size. These networks
have the highest mean degree distribution (mean = 84,
std = 132), but the lowest average clustering coefficient.
The networks are scattered in many separate connected
components, with the biggest covering 2% of the network,
and thus show the highest modularity score. The game
networks in this cluster also have the highest degree
assortativity, and most of them (88%) are scale-free
networks. Finally, they are relatively centralized, although
not as much as the networks in Cluster 0.
• Cluster 3. In this cluster, we have 13 games. The net-
works are, on average, much bigger than in Cluster 0 and
2. The degree distribution has a quite high mean, although
being very variable (mean = 74, and std = 271). The
networks show a high modularity and are divided into
many components, with the largest accounting for 18%
of the nodes. Finally, the networks are not particularly
centralized, and only about half of them are scale-free.
• Cluster 4. This cluster includes one title: Wallpaper En-
gine, whose network counts about 5k nodes. The degree
distribution is quite low, but very variable (mean = 27,
and std = 214), and does not follow the power-law
distribution. The network shows the lowest modularity
score and is not particularly centralized.
• Cluster 5. This cluster includes only Counter-Strike:
Global Offense (CS:GO), the network of which is much
bigger than the others (17k nodes). The network has the
highest clustering coefficient and a big LCC (71% of
the network). The CS:GO network is also a scale-free
network, with the lowest centralization score.
In addition to inspecting the network features, we also
studied how the tags defined by the users in Steam are
distributed in each cluster. In order to do this, we build a word
cloud for each cluster (Figure 2). The clouds are obtained
through the wordcloud python package, and compute the
word (i.e., tag) frequency in the text (i.e., cluster tag list).
Cluster 1, 4, and 5 include only one game each, and thus the
tag displayed entirely describe that specific game title. Rocket
League (Cluster 1) is a team-based, competitive game,
which often is used with the local multiplayer feature to build
the teams. Wallpaper Engine (Cluster 4) is a modeling
and design tool to make animated wallpaper for windows.
Finally, CS:GO (Cluster 5) is a team-based competitive
game, which falls under the category of esport. On the other
TABLE III: Summary of the clusters and their characterization — the percentages in brackets indicate how frequent the game
type was in the cluster, for the clusters with more than one element.
Cluster Size Game type (from user-defined tags) Game network
#0 106 Simulation (100%), online coop (70%), casual (70%) Dense, low clustering coeff, centralized structure (degree)
#1 1
Local coop, local multiplayer, team-based (RL)
Scale-free, low degree, centralized structure (betweenneess), LCC 50%
#2 72
Massively multiplayer (100%), F2P (90%),
singleplayer (73%)
88% scale-free, scattered, high modularity, low clustering coeff, small networks
#3 13 Online coop (100%), team-based (70%) High modularity, LCC 18%, quite big networks
#4 1
Not a game but a design tool (Wallpaper Engine)
5k players, but low degree, lowest modularity
#5 1
Esport, team-based, competitive (CS:GO)
17k players, scale-free, highest clustering coeff, LCC 71%, not centralized
hand, the other word clouds are more crowded. Cluster
0 has mostly cooperative, casual games, which are in third
person and/or simulation. In Cluster 2, the differences are
less delineated. Nevertheless, the games are mostly free-to-
play, with a massively multiplayer and single-player mechanic
combined. Finally, Cluster 3 is characterized by team-
based games, both cooperative and competitive, with elements
of crafting and building.
The game genres are uniformly distributed across the clus-
ters. Except for Cluster 1, 4, and 5, whose genres reflect
the single game that they represent, the other clusters lack
a narrow distinction of the characterizing genres. Specifically,
the most popular genre was Action (17% for Cluster 0, 13%
for Cluster 2, and 30% for Cluster 3), followed by Adventure
(13% for Cluster 0, 12% for Cluster 2, and 9% for Cluster 3),
Simulation (12% for Cluster 0, 10% for Cluster 2, and 15% for
Cluster 3), and RPG (13% for Cluster 0, 8% for Cluster 2, and
8% for Cluster 3). The genre property thus does not contribute
to the cluster characterization as much as user-defined tags did.
Table III summarizes these findings.
VI. DISCUSSION
In this work, we analyzed the Steam friendship graph for
200 games to connect their network properties to their charac-
teristics (i.e., genre and user-defined tags). Our investigation
led to six clusters, summarized in Table III.
We observed how user-defined tags were more informative
than game genres, as the genres are distributed in similar
ways across the six clusters. Although we should take into
consideration the fact that the user-defined tags also go down
to a greater level of detail, this result suggests how the way
players play games is related to the structure of the commu-
nity. The game network analyses can thus provide valuable
insights into how the game is perceived and interpreted by its
community and its dominant social dynamics. In other words,
the players are essential for shaping the gaming experience
and its mechanics on a level similar to that of the designers.
This finding also emphasized the importance of analyzing the
player network not only to detect interaction patterns, but also
to some extent for obtaining design feedback.
The cluster characterizations also provided more in-depth
correlations among the community structures and the game
properties. First, we observed that team-based games tended
to have scale-free networks, regardless of the network sizes
(Cluster 1 and 5). In other words, the degree distributions of
their nodes follow a power-law distribution, in which many
nodes have few connections but only a few nodes have a large
number of links. This is also known as the richer-get-richer
phenomenon, prevalent in social media networks. Therefore,
the network structure of games with a strong team component
shows some similarities to social media communities [36],
[37], suggesting that playing in teams creates an environment
in which social relationships are nurtured. The existence of
teams in itself also impacted the clustering coefficient, which
was higher than in cooperative but casual games (Cluster 0).
The result here is that playing in teams increases the likelihood
of becoming friends of friends in Steam than sharing a
similar game preference (either cooperative or single-player).
Additionally, multiplayer games without a strong team-based
component (Cluster 0 and 2) tended to have much more
centralized networks than games in which teams are important
(Cluster 1, 3, and 5). Hence, having players’ connected in
teams uniformed the link distribution, increasing the likelihood
of forming triangles (i.e., friends of friends).
Most of the massively multiplayer games with a single-
player component (Cluster 2) also showed a tendentially scale-
free network structure. These networks were highly scattered
in small graphs with high modularity, however, whereas team-
based game graphs were more connected. This finding further
strengthens the importance of teams or team tasks to obtain a
more cohesive community, especially in terms of relationships
beyond the game (i.e., a Steam friendship). Furthermore, the
scattered nature of these networks can be linked to them being
massively multiplayer games. The result of this is that players
come into contact with many (random) users. Consequently, it
is unlikely that they will form a (Steam) friendship or play
in more matches together. The graph of Cluster 4, which
represents a design tool only usable in a single-user mode,
is similarly scattered and sparse. This characterization lends
support to the importance of an in-game social aspect for
nurturing the game community.
Finally, the results show that the biggest network with the
highest clustering coefficient is a team-based esport game
(CS:GO). Games of this type rely on the idea of (consolidated)
teams more than others and do so not only for professionals but
also for amateurs. This might contribute to the high clustering
coefficient, as repeatedly playing together might result in a
Steam friendship (or vice-versa friends form stable teams).
In summary, the shape of the game community is more
affected by how players define the game (through user-
defined tags) than by the game genre assigned by the design-
ers/producers. Specifically, games that require the formation of
teams show specific network properties: they are not central-
ized, have a high clustering coefficient, and tend to be scale-
free. They support the formation of friendships to a greater
extent than non-team-based games. Conversely, titles with a
stronger single-player focus have more scattered networks,
made of many small detached components.
Limitations
This study also comes with a few limitations. First, our
population sample comprehends a subset of the actual Steam
network. Consequently, we have a partial view of the friends
graph and the network in each game. Additionally, our sample
is seeded from the players reviewing “new and trending
games” and playing at least one game within our observation
period. This means we must interpret our results in terms of
active players (in our 5-week time window). Furthermore, the
resulting friendship networks may also have been partially
influenced by the recent pandemic since more people may
have been obliged to stay at home and will have had more
time to play games during this period. Second, the game
networks vary in size and the sub-graph structure might not
entirely reflect the full-game network, in particular for the
smaller networks. Since some network-related features could
have accentuated this bias, we relied on graph embeddings
to perform the comparison. Finally, we do not consider the
temporal component in our study. Hence, we cannot infer
causality in the relationships between graph topology and
friendship connections. In other words, we can neither argue
that players became friends because they played the same
game title or that they began playing because they were already
friends. Rather, we correlate the Steam community structure of
each game to the game properties to identify patterns linking
the network properties and game type (i.e., user tags).
Future Works
The natural next step for our study will be the introduction
of the temporal factor in the analysis. In this study, we cannot
infer causality among the two variables studied: the game
network structure and the game tags. In future work, we
aim to achieve an understanding of the link between game
tags and how they can be used to predict new games, and
vice-versa (i.e., the link between game tags and how they
can be used to predict new friends). Additionally, we will
analyze the network evolution for each game to understand the
dynamics by which the connections are formed (e.g., velocity
and patterns). Finally, we will investigate how we can infer
the (used) game mechanics and playstyle from the network
structure. In this regard, we will verify how users play the
game despite the available mechanics. For example, a game
might be both cooperative and competitive, but the network
might prefer building a community through cooperation.
VII. CONCLUSION
People connect through games, in which they build rela-
tionships and friendships that can be translated into other
games and also into real world situations. Recent research on
the player networks induced from in-game social interactions
showed the potential of SNA in games towards unraveling the
dynamic occurring within the game community. Nevertheless,
a broader understanding of how those game networks are
shaped and formed still lacks. Also lacking is an analysis of
how the game characteristics are related to the structure of
its player network. This study deepened our understanding of
player communities by analyzing the Steam friendship graphs
for 200 games. We found that some game characteristics are
linked to specific network properties. The game community’s
shape is related to the tags users associated with games,
more than the game official genres. Specifically, team-based
games have scale-free networks that are not centralized and
with a high clustering coefficients. Hence, they support the
formation of friendships to a far greater extent than non-team-
based games. On the other hand, single-player games have
scattered networks made of many small detached components.
In conclusion, the type of game is indeed reflected in how
the community is shaped and formed, supporting the idea that
playing together (especially in teams) is more likely to produce
social capital, and thus (Steam) friendships.
REFERENCES
[1] N. Johannes, M. Vuorre, and A. K. Przybylski, “Video game play
is positively correlated with well-being,” Royal Society Open Science,
vol. 8, no. 2, p. 202049, 2021.
[2] A. K. Przybylski, C. S. Rigby, and R. M. Ryan, “A Motivational Model
of Video Game Engagement,” Review of General Psychology, vol. 14,
no. 2, pp. 154–166, 2010.
[3] J. Pirker, A. Rattinger, A. Drachen, and R. Sifa, “Analyzing player
networks in Destiny,” Entertainment Computing, vol. 25, pp. 71–83,
2018.
[4] M. H. Schiller, G. Wallner, C. Schinnerl, A. M. Calvo, J. Pirker, R. Sifa,
and A. Drachen, “Inside the group: Investigating social structures in
player groups and their influence on activity,” IEEE Transactions on
Games, vol. 11, no. 4, pp. 416–425, 2018.
[5] R. L. Mandryk, J. Frommel, A. Armstrong, and D. Johnson, “How
Passion for Playing World of Warcraft Predicts In-Game Social Capital,
Loneliness, and Wellbeing,” Frontiers in Psychology, vol. 11, p. 2165,
2020.
[6] L. E. Nacke, M. Klauser, and P. Prescod, “Social Player Analytics in a
Facebook Health Game,” in Proceedings of KHCI, 2015, pp. 180–187.
[7] G. Wallner, S. Kriglstein, and A. Drachen, “Tweeting your destiny:
Profiling users in the twitter landscape around an online game,” in 2019
IEEE Conference on Games (CoG). IEEE, 2019, pp. 1–8.
[8] G. Wallner, C. Schinnerl, M. H. Schiller, A. Monte Calvo, J. Pirker,
R. Sifa, and A. Drachen, “Beyond the individual: Understanding social
structures of an online player matchmaking website,” Entertainment
Computing, vol. 30, p. 100284, 2019.
[9] Y. J. Kim, D. Engel, A. W. Woolley, J. Y. T. Lin, N. McArthur, and T. W.
Malone, “What makes a strong team? Using collective intelligence to
predict team performance in League of Legends,” in Proceedings of
CSCW, 2017, pp. 2316–2329.
[10] E. Loria, J. Pirker, A. Drachen, and A. Marconi, “Do influencers
influence?–analyzing players’ activity in an online multiplayer game,”
in International Conference on Games (CoG). ACM, 2020.
[11] R. F. Baumeister and M. R. Leary, “The Need to Belong: Desire
for Interpersonal Attachments as a Fundamental Human Motivation,”
Psychological Bulletin, vol. 117, no. 3, pp. 497–529, 1995.
[12] R. Rogers, “The motivational pull of video game feedback, rules,
and social interaction: Another self-determination theory approach,”
Computers in Human Behavior, vol. 73, pp. 446–450, 2017.
[13] A. Voida, S. Carpendale, and S. Greenberg, “The individual and the
group in console gaming,” in Proceedings of CSCW, 2010, pp. 371–
380.
[14] N. Crenshaw and B. Nardi, “What’s in a name? Naming practices in
online video games,” in Proceedings of CHI Play, 2014, pp. 67–76.
[15] J. S. Saltz, S. R. Hiltz, and M. Turoff, “Student social graphs,” in
Proceedings of CSCW. New York, New York, USA: ACM (ACM),
2004, p. 596.
[16] G. Freeman, “Making games as collaborative social experiences: Ex-
ploring an online gaming community,” in Proceedings of CSCW, vol.
26-February-2016, 2016, pp. 265–268.
[17] N. Ducheneaut, R. J. Moore, and E. Nickell, “Virtual ”third places”:
A case study of sociability in massively multiplayer games,” Computer
Supported Cooperative Work, vol. 16, no. 1-2, pp. 129–166, 2007.
[18] N. Ducheneaut and R. J. Moore, “The Social Side of Gaming: A Study
of Interaction Patterns in a Massively Multiplayer Online Game,” in
Proceedings of CSCW, 2004.
[19] Y. Xu, X. Cao, A. Sellen, R. Herbrich, and T. Graepel, “Sociable killers:
Understanding social relationships in an online first-person shooter
game,” in Proceedings of CSCW, 2011, pp. 197–206.
[20] Y. Huang, W. Ye, N. Bennett, and N. S. Contractor, “Functional or
social? Exploring teams in online games,” in Proceedings of CSCW.
New York, New York, USA: ACM Press, 2013, pp. 399–408.
[21] J. Jiang and S. Yarosh, “Why do teammates hate me? Cross-cultural
tensions and social dynamics in online games,” in Proceedings of CSCW,
vol. 26-February-2016, 2016, pp. 301–304.
[22] C. S. Ang and P. Zaphiris, “Social Roles of Players in MMORPG
Guilds,” Information, Communication & Society, vol. 13, no. 4, pp. 592–
614, 2010.
[23] N. Ducheneaut, N. Yee, E. Nickell, and R. J. Moore, “”Alone To-
gether?” Exploring the Social Dynamics of Massively Multiplayer
Online Games,” in Proceedings of CHI. New York, New York, USA:
ACM Press, 2006.
[24] H. Park and K. J. Kim, “Social network analysis of high-level players
in multiplayer online battle arena game,” in Lecture Notes in Computer
Science, vol. 8852. Springer Verlag, 2015, pp. 223–226.
[25] A. Rattinger, G. Wallner, A. Drachen, J. Pirker, and R. Sifa, “Integrating
and inspecting combined behavioral profiling and social network models
in Destiny,” in Lecture Notes in Computer Science, vol. 9926 LNCS.
Springer Verlag, 2016, pp. 77–89.
[26] Z. Zuo, “Sentiment Analysis of Steam Review Datasets using Naive
Bayes and Decision Tree Classifier,” Tech. Rep., 2018.
[27] R. Sifa, A. Drachen, and C. Bauckhage, “Large-Scale Cross-Game
Player Behavior Analysis on Steam,” Tech. Rep., 2015.
[28] R. Becker, Y. Chernihov, Y. Shavitt, and N. Zilberman, “An analysis of
the Steam community network evolution,” 2012 IEEE 27th Convention
of Electrical and Electronics Engineers in Israel, IEEEI 2012, 2012.
[29] M. O’Neill, E. Vaziripour, J. Wu, and D. Zappala, “Condensing steam:
Distilling the diversity of gamer behavior,” Proceedings of the ACM
SIGCOMM Internet Measurement Conference, vol. 14-16-Nove, pp. 81–
95, 2016.
[30] H. Cai, V. W. Zheng, and K. Chang, “A comprehensive survey of graph
embedding: Problems, techniques, and applications,” IEEE Transactions
on Knowledge & Data Engineering, vol. 30, no. 09, pp. 1616–1637,
2018.
[31] A. Narayanan, M. Chandramohan, R. Venkatesan, L. Chen, Y. Liu, and
S. Jaiswal, “graph2vec: Learning distributed representations of graphs,”
in Proceedings of the 13th International Workshop on Mining and
Learning with Graphs, 2017.
[32] T. Mikolov, K. Chen, G. Corrado, and J. Dean, “Efficient estimation of
word representations in vector space,” in 1st International Conference
on Learning Representations, Workshop Track Proceedings, 2013.
[33] P. Rousseeuw, “Silhouettes: A graphical aid to the interpretation and
validation of cluster analysis,” Journal of Computational and Applied
Mathematics, vol. 20, pp. 53–65, 1987.
[34] J. Overgoor, B. State, and L. A. Adamic, “The structure of u.s.
college networks on facebook,” Proceedings of the International AAAI
Conference on Web and Social Media, vol. 14, no. 1, pp. 499–510, 2020.
[35] C. D. Manning, P. Raghavan, and H. Sch
¨
utze, Introduction to Informa-
tion Retrieval. Cambridge University Press, 2008.
[36] A. Mislove, M. Marcon, K. P. Gummadi, P. Druschel, and B. Bhattachar-
jee, “Measurement and analysis of online social networks,” in Proceed-
ings of the 7th ACM SIGCOMM conference on Internet measurement,
2007, pp. 29–42.
[37] A.-L. Barab
´
asi, Network Science. Cambridge University Press, 2016.
APPENDIX
A. List of the Steam titles contained in each cluster.
Cluster Steam Titles
#0 Counter-Strike, Counter-Strike: Source, Left 4 Dead 2, Portal
2, Sid Meier’s Civilization V, Grand Theft Auto IV: The
Complete Edition, Total War: Shogun 2, Red Dead Redemp-
tion 2, FinalFantasy XIV Online, Mount & Blade: Warband,
Borderlands 2, The Elder Scrolls V: Skyrim, Tomb Raider,
Alien: Isolation, Payday 2, Grim Dawn, DayZ, Insurgency,
Ghostrunner, Euro Truck Simulator 2, Wreckfest, Warframe,
Company of Heroes 2,Killing Floor 2, War Thunder, Star
Wars Jedi: Fallen Order, Europa Universalis IV, Path of Exile,
Dying Light, The Forest, Assetto Corsa, Plague Inc: Evolved,
The Binding of Isaac: Rebirth, Cities: Skylines, Tesla Effect:
A Tex Murphy Adventure, Mount & Blade II: Bannerlord,
Destiny 2, Fallout 76, XCOM 2, American Truck Simulator,
No Man’s Sky, Dota Underlords, Stellaris, 100% Orange Juice,
Sid Meier’s Civilization VI, Brawlhalla, The Witcher 3: Wild
Hunt, Unturned, The Elder Scrolls Online, Call of Duty: Black
Ops III, Geometry Dash, Don’t Starve Together, ARK: Sur-
vival Evolved, Black Mesa, Dark Souls III, Human Resource
Machine, Fallout 4, Doom, Scrap Mechanic, Tekken 7, Rise
of the Tomb Raider, Squad, Hearts of Iron IV, Tower Unite,
Borderlands 3, Business Tour - Board Game with Online
Multiplayer, Stardew Valley, Factorio, Golf With Your Friends,
VRChat, Friday the 13th: The Game, Pac-Man Championship
Edition 2, Pinball FX3, Paladins, UNO, Cyrano Story, Human:
Fall Flat, Minion Masters, The Elder Scrolls V: Skyrim Special
Edition, Manual Samuel - Anniversary Edition,Ashes of the
Singularity: Escalation, Redout: Enhanced Edition, Deep Rock
Galactic, Warhammer: Vermintide 2, Playerunknown’s Bat-
telgrounds, Insurgency: Sandstorm, Assassin’s Creed Origins,
Total War: Warhammer, Streets of Rage 4, Hunt: Showdown,
Mordhau, Slay the Spire, Jurassic World Evolution, Generation
Zero, Shadow of the Tomb Raider: Definitive Edition, Project
Winter, Doom Eternal, Farming Simulator 19, Assassin’s
Creed Odyssey, Age of Empires II: Definitive Edition, Green
Hell, HITMAN 2, Pummel Party, XCOM: Chimera Squad,
Resident Evil 2, Resident Evil 3, Mortal Kombat 11, Halo:
The Master Chief Collection
#1 Rocket League
#2 Half-Life, Half-Life 2, Star Wars: Battlefront 2, Call of Duty:
World at War, Grand Theft Auto: San Andreas, Fallout:
New Vegas, Call of Duty: Black Ops, Realm of the Mad
God Exalt, Call of Duty: Black Ops II, Crusader Kings II,
SpeedRunners, Starbound, Bloons TD 6 PlanetSide 2, Age of
Empires II, Just Cause 3, Deadside, Heroes & Generals, Prison
Architect, Space Engineers, 7 Days to Die, Darkest Dungeon,
Subnautica, tModLoader, BeamNG.drive, Metal Gear Solid V:
The Phanthom Pain, Warface, RimWorld, Trove, For Honor,
DiRT Rally, Enter the Gungeon, Town of Salem, Dark Souls II:
Scholar of the First Sin, Elite Dangerous, Astroneer, Blender,
Hollow Knight, Project Cars 2, Smite, CRSED: F.O.A.D., The
Jackbox Party Pack 2, Rising Storm 2: Vietnam, Blackwake,
The Jackbox Party Pack 3, KovaaK 2.0, F12019, Divinity:
Original Sin 2 - Definitive Edition, Conan Exiles, Drawful
2, Deceit, Northgard, Planet Coaster, Half-Life: Alyx, Far
Cry 5, Dark Souls: Remastered, Black Desert Online, Dead
Cells, The Jackbox Party Pack 4, House Flipper, Remnant:
From the Ashes, Beat Saber, Soundpad, Raft, Stick Fight: The
Game, Dragon Ball FighterZ, DiRT Rally 2.0, SCP: Secret
Laboratory, Aim Lab, Post Scriptum, Albion Online, Sekiro:
Shadows Die Twice - Goty Edition
#3 Team Fortress 2, Dota 2, Garry’s Mod, Terraria, Arma 3,
SteamVR, Rust, Grand Theft Auto V, Tabletop Simulator,
Tom Clancy’s Rainbow Six Siege, Dead by Daylight, Monster
Hunter: World, Risk of Rain 2
#4 Wallpaper Engine
#5 Counter-Strike: Global Offensive
