Application of Community Detection Algorithms on

Social Internet-of-things Networks

Abstract—The Internet-of-things (IoT) networks are witnessing diverse domains such as smart cities, smart grid, and crowd-
a drastic increase over the years. Twenty billion devices connected sourcing, to name a few. For instance, for the latter application,
to the Internet are expected in 2022. The need for identifying com- based on their social relationships, a crowdsourcer can recruit
munities within such networks can serve as a strong complexity
reduction mean for many discovery and identification services. a set of trustworthy and socially connected objects that will
The idea of communities in IoT networks is also motivated act as workers to execute its request.
by the emerging concept of socializing IoT devices. In this Most of the real-world networks, e.g., social networks, are
paper, we investigate the application of two community detection not random graphs. On the contrary, they can reveal a high
algorithms, namely Louvain and Bron-Kerbosch algorithms, on
IoT networks usually represented by large-scale graphs. The
level of order and organization. Usually, the degree distribution
objective is to convert the complex IoT network into multiple follows a power law: therefore, a low number of vertices have
overlapping and non-overlapping communities where its elements a high degree number while the rest have a small degree
share common characteristics. Starting from a real-world IoT [5]. These properties are also observed in real-world SIoT
networks, we use its dataset to extract community-structured networks, which motivates this work to design community
IoT network based on different types of relationships such as
co-location, owner social relationships, and autonomously build
detection algorithms to determine communities within SIoT
object relationships among objects. Our analysis showcases how devices. More motivations raise with the issues and needs
community detection algorithms structure the IoT network into for SIoT service discovery and composition, and network
communities based on the different relationships established navigation. Additional motivation emerges with the concept
between objects. of mobile IoT crowdsourcing applications [6] where it is rec-
Index Terms—Internet of things (IoT), community detection,
clustering.
ommended to reduce the complexity of the recruitment process
by limiting the IoT workers’ search space. Hence, instead of
I. I NTRODUCTION
looking at the whole network, selected IoT communities with
Internet-of-things (IoT) is a vast ubiquitous network of specific characteristics can constitute a smaller search space
connected heterogeneous devices. The number of devices con- for crowdsourcing platforms.
nected to the Internet could reach 29 billion by 2022 according In this study, we intend to investigate community detection
to Ericsson [1]. These devices range from simple sensors algorithms for social IoT networks. The objective is to or-
such as weather and humidity sensors to highly computational ganize IoT networks and gather IoT devices in communities
devices such as smartphones [2]. This makes the network and sharing mutual interest and characteristics. We proceed with
resource management of such large-scale systems complicated community detection approaches applied on different relation-
and challenging for both researchers and industry. ship graphs, e.g., co-located and/or socially connected devices,
Each node in the IoT network can be a standalone automated using overlapping and disjoint algorithms, i.e. Louvain and
device or human-controlled entity connected via various com- Bron-Kerbosch algorithms. Then, we analyze the obtained
munication technologies such as cellular networks or device- communities of a real-world IoT network and investigate the
to-device communications and establishing human-to-human, performance of the algorithms.
human-to-machine, and machine-to-machine links. The rela-
tionships between IoT devices are not purely for networking II. M ETHODOLOGY
and data relaying. Nowadays, IoT objects can also have social
relationships allowing them interacting with each other for The proposed framework, presented in Fig. 1, starts from a
diverse purposes, e.g., share resources such as memory and dataset of IoT devices located in a vast geographical area. The
CPUs or exchange information in order to improve the quality dataset includes different types of relations represented in the
of service (QoS) and quality of experience (QoE) of users from form of graphs and adjacency matrices. From these relations,
an application perspective. The concept of social Internet- we aim to perform community detecting algorithms to extract
of-things (SIoT) is then introduced as a means to organize overlapping and disjoint communities, which will be used to
and manage the social relationships among interacting objects cluster IoT devices into groups sharing common characteristics
given mutual interest and trustworthiness [3], [4]. This concept and also assist in identifying specific devices to facilitate their
helps in enhancing the operation of several applications in search process in future applications.

978-1-7281-4058-2/19/$31.00 ©2019 IEEE

94
 Fig. 1: Proposed framework for community detection in large-scale social IoT networks.
A. Dataset • Social network device relation (SNDR) Social objects re-
We are using a dataset provided by Marche et al. in [7]. lations can be based on the social relationships of their owners.
The dataset includes real IoT objects in the city of Santander, For example, if the owners of these devices are “friends” then
Spain, mixed with simulated objects such as smartphones, the devices can establish relationships accordingly.
tablets, and personal computer devices. The total number of • Social object relation (SOR): this relationship is au-
objects is 16216 devices, of which 14600 from private users tonomously established when the objects come into contact.
and 1616 from public services. Fig. 2 shows the available Their exchange can be sporadically or continuously based
information about each device in the dataset, such as the type, on the needs of the devices and/or the owner policies. The
the brand, and the model. Also, the dataset indicates whether relationship can be then established based on their previous
a device is static or mobile and provides their geographical experiences.
locations (x and y). For private devices, additional information The four different relations mentioned earlier are converted
about the owner identification is stored in user id. into four graphs. All of them are unweighted and undirected.
Also, they have no self-loop links. By keeping that in mind,
in the next section, we choose two algorithms to detect
communities of this IoT network applicable to this structure
of graphs.
III. C OMMUNITY D ETECTION A LGORITHMS
The community detection algorithms are used to automat-
ically extract relevant groups of nodes and label them as
communities. The concept of a community structure dated
back to 1962 H.A. Simon. There are different applications
of community detection methods, such as understanding the
structure of complex systems such as large-scale IoT networks.
Also, it helps in visualizing these systems by labeling and
Fig. 2: High-level entity relationship of the SIoT Dataset from clustering the nodes of the graph. Moreover, it can be used
Marche et al. [7]. to improve information retrieval, in our case, by helping in
For the relationships between the devices, the definition service and device discovery.
of links can impact the network navigability and service In Fig. 3, we provide a basic example explaining the differ-
discovery. The types of relationships between objects can ence between disjoint and overlapping communities. The node
be established based on the historical usage and context of in the middle belongs to a unique cluster which is community
the SIoT devices. Generally, there are four types of relations 3 in a disjoint context, while, in the overlapping communities,
between the objects in the SIoT context that are proposed by the middle node is associated to two communities 1 and 2.
Atzori et al. [8]. The classes are listed below: Based on that, we can divide the community detection into
• Co-location/co-work based relation (CLOR): the spatial two types based on their ability to capture overlapping nodes
properties of the device influence the network structure. For under multiple communities. The Louvain algorithm to assign
example, if a device shares a physical location (co-habitation) each node to only one community [9] and the Bron-Kerbosch
or is within the same network domain of another device, then algorithm [10] to determine overlapping communities.
we create the relationship between these devices.
• Object ownership relation (OOR): is determined based on A. Louvain Algorithm
the ownership of the object. This network can be influenced The Louvain algorithm is proposed in 2008 to detect
on how the objects are connected to each other. As a simplest disjoint communities in a graph. The technique maximizes
example, the ownership of set of smart devices can influence the modularity score for each community, which represents
the things connectivity and type of authentication needed the quality of assigning nodes to communities by examining
to access them. The devices such as phones, tablets, game the density of edges with a set of nodes compared to how
console, home sensors, etc. can be connected based on the it would be connected in a random network. The Louvain
ownership of these devices. method is one of the fast community detection algorithms. It

95
 the elements of each community. When the second phase is
finished, the algorithm returns to the first phase and applies the
two stages process at a higher-level graph produced from the
previous iteration (the vertices are replaced by communities).
This process continues until the modularity score reaches the
maximum possible score for each node in the graph. Hence,
each node will belong to one community.
B. Bron-Kerbosch Algorithm
The Bron-Kerbosch algorithm [10], aka maximal clique, is a
polynomial algorithm of complexity O(3n3 ). The algorithm is
recursive backtracking, which enumerates all maximal cliques
by searching all the complete connected nodes within the
Fig. 3: An illustration to differentiate between disjoint and network. The Bron-Kerbosch is suitable for an unweighted
overlapping communities. undirected graph to find overlapping communities. The dataset
we are using in this study demonstrates these topological
is a greedy optimization method with a complexity O(n log n). properties. Indeed, finding overlapping communities remains
It maximizes the modularity value, which is between 1 and in our interest since the devices can belong to multiple
1. There are multiple modularity measures in the literature. communities in real-world applications, especially when we
Despite that, in [9], the modularity measures the density of consider SNDR and SOR relations.
links inside communities. The Louvain algorithm uses the
following modularity score for a weighted graph: Algorithm 1 Bron-Kerbosch Algorithm
 1: procedure B RON K ERBOSCH 1(R, P, X)
1 X ki kj
Q= Aij (ci , cj ), (1) 2: if P and X are empty then
2m ij 2m
3: R maximal clique
where: 4: for vertex v in P do
• Aij the weight of edge between nodes i and j; 5: BronKerbosch1(
S
• ki and kj are the sum of the weights of the edges attached 6: R T {v},
to nodes i and j, respectively; 7: P T N (v),
• 2m is the sum of all of the edge weights in the graph; 8: X N (v) )
• ci and cj are the communities of the nodes; and 9: P := P \S{v}
• is a simple Kronecker delta function. 10: X := X {v}
The method has two stages to maximize the modularity
score efficiently. In the first phase, each node in the network Based on [11], initially, the Bron-Kerbosch Algorithm 1,
is labeled in its own community. Then, for each node i, the starts with empty R and X sets where R handles temporary
change of modularity is calculated by removing the node i results and X denotes the excluded set. The set P is the set
from its community and moving it into the neighbor commu- of possible candidates, while N (v) stores the neighbors of the
nity. The difference in the modularity score is calculated in node v. The algorithm starts by picking one node v from the
two steps. Removing i from its community and inserting it in set P . Next, it adds the selected v to R set and removes it from
the community of its neighbor node j. In the second phase, the its non-neighbor nodes from P and X. Then, it picks another
following equation is calculated to find the difference between node from the set P . The process is repeated until there are no
the modularity of the nodes and their neighbors: further nodes in P set. When P and X are again empty then set
 ✓ ◆2 R represents the new maximal clique. Otherwise, R contains
⌃in + 2ki,in ⌃tot + ki a clique in the subset. After that, the algorithm backtracks last
Q=
2m 2m picked node and stores the content of P , R, and X before the
 ✓ ◆2 ✓ ◆2
⌃in ⌃tot ki node process selection. It then removes the node from P and
(2) adds it to X, then expands to the next node. If P is ; then
2m 2m 2m
the algorithm backtracks to the higher-level.
The second phase of the algorithm groups all the nodes
within the same community into one node as a higher-level IV. R ESULTS & D ISCUSSION
graph from the previous stage. The edges of the same commu- The devices are projected into a xy-plane between 0 and
nity will be presented as a self-loop to the node that represents 1. We pick a subset of the devices between an area of four
the community whose number of edges serves the weight of vertices between 0.6 to 0.8. The selection resulted in 2,567
the self-loop. The multiple connections of the community with devices from the 16,216 to do a pilot study on the dataset.
other communities are reduced to one edge with a weight The small subset of the objects will help us to understand
representing the number of connections of its elements with analyze both algorithms before conducting further analysis

96
 (a) CLOR Louvain (b) SOR Louvain (c) SNDR Louvain
Smartphones Environment Sensors Home Sensors Indicators Parking Sensors PCs
Printers Fitness Devices Car Streetlight Sensors Alarms Tablet
Transportation Smartwatch Others

Fig. 4: The frequency of objects for each community detected using Louvain algorithm.
TABLE I: Obtained communities from Louvain and Bron- IoT network into communities-structured network to ease its
Kerbosch algorithms on randomly selected IoT devices discovery and object identification for different services and
Louvain Method Bron-Kerbosch Method applications. Both disjoint and overlapping communities are
# of Comm. Avg. Size # of Comm. Avg. Size
detected using Louvain and Bron-Kerbosch algorithms, respec-
CLOR 1787 1.44 3017 1.09 tively. We have applied the community detection algorithms on
SOR 2259 1.14 2602 1.16
SNDR 164 15.65 7886 3.08 a real-world dataset and visualized the obtained communities
from different relationship criteria. As a future work, we will
on the remaining devices due to computational limitation. consider the mobility of IoT devices and designing dynamic
We run the two algorithms, Louvain and Bron-Kerbosch, on communities detection approaches.
three adjacency matrices representing the subset of the devices R EFERENCES
where each matrix’s element indicates an edge based on the
[1] Ericsson - Internet of Things forecast, 2019. Available at:
previous definitions of CLOR, SOR, and SNDR relationships. https://www.ericsson.com/en/mobility-report/internet-of-things-forecast
Fig. 4 demonstrates the number of nodes that belongs Accessed: 2019-08-27.
to each community in Louvain. For each algorithm, we set [2] A. Bader, H. Ghazzai, A. Kadri, and M. Alouini, “Front-end intelligence
for large-scale application-oriented internet-of-things,” IEEE Access,
a threshold for visualization purposes. All the communities vol. 4, pp. 3257–3272, June 2016.
having less than four devices are plotted under the “others” bin [3] M. Kranz, L. Roalter, and F. Michahelles, “Things that twitter: social
in Fig. We didn’t visualzie Bron-Kerbosch method since the networks and the internet of things,” in What can the Internet of
Things do for the Citizen (CIoT) Workshop at The Eighth International
communities are small in size and all will be in “others” bins. Conference on Pervasive Computing (Pervasive’10), Helsinki, Finland,
Besides that, each color in the bar graph shows a different type May 2010.
of devices listed in the legend of Fig. 4. Mostly, for different [4] B. Afzal, M. Umair, G. A. Shah, and E. Ahmed, “Enabling IoT platforms
for social IoT applications: vision, feature mapping, and challenges,”
relations, we find that each community contains multiple types Future Generation Computer Systems, vol. 92, pp. 718–731, Mar. 2019.
of devices with no specific patterns. But, in SDNR, we observe [5] S. Fortunato, “Community detection in graphs,” Physics reports,
that one type of devices in each community are dominating vol. 486, no. 3-5, pp. 75–174, Feb. 2010.
[6] A. Hamrouni, H. Ghazzai, M. Frikha, and Y. Massoud, “A photo-based
the community. mobile crowdsourcing framework for event reporting,” in IEEE Int.
In Table I, the number of communities found by each Midwest Symp. Circuits Syst., Dallas, TX, USA, Aug. 2019.
method and the average size of each community are provided. [7] C. Marche, L. Atzori, and M. Nitti, “A dataset for performance anal-
ysis of the social internet of things,” in IEEE Annual International
We notice that the average size of each community are small Symposium on Personal, Indoor and Mobile Radio Communications
except for SNDR communities. We examined five devices (PIMRC’18), IEEE, Bologna, Italy, Sept. 2018.
from the results to investigate their respective communities. [8] L. Atzori, A. Iera, G. Morabito, and M. Nitti, “The social internet of
things (siot)–when social networks meet the internet of things: Concept,
We notice that devices can be significantly affected based on architecture and network characterization,” Computer networks, vol. 56,
the definition of the link. For example, we notice that relations no. 16, pp. 3594–3608, Nov. 2012.
of Smartphone, Car, and Fitness where the SNDR relation [9] V. D. Blondel, J.-L. Guillaume, R. Lambiotte, and E. Lefebvre, “Fast
unfolding of communities in large networks,” Journal of statistical
using the Louvain method but their co-location communities mechanics: theory and experiment, vol. 2008, no. 10, Oct. 2008.
labels are diverse. [10] C. Bron and J. Kerbosch, “Algorithm 457: finding all cliques of an
undirected graph,” Communications of the ACM, vol. 16, no. 9, pp. 575–
V. C ONCLUSION 577, Sept. 1973.
[11] A. Conte, “Review of the bron-kerbosch algorithm and variations,”
In this study, we proposed a framework to detect commu- School of Computing Science, University of Glasgow, pp. 1–9, May
nities of social devices in IoT networks using two community 2013.
detection algorithms. The objective is to convert the complex

97