Louvain-like Methods for Community Detection in Multi-Layer

Networks
Sara Venturini∗, Andrea Cristofari†, Francesco Rinaldi‡, Francesco Tudisco§

Abstract. In many complex systems, entities interact with each other through complicated patterns that
embed different relationships, thus generating networks with multiple levels and/or multiple types of edges.
arXiv:2106.13543v1 [cs.SI] 25 Jun 2021

When trying to improve our understanding of those complex networks, it is of paramount importance to
explicitly take the multiple layers of connectivity into account in the analysis. In this paper, we focus on
detecting community structures in multi-layer networks, i.e., detecting groups of well-connected nodes
shared among the layers, a very popular task that poses a lot of interesting questions and challenges. Most
of the available algorithms in this context either reduce multi-layer networks to a single-layer network or
try to extend algorithms for single-layer networks by using consensus clustering. Those approaches have
anyway been criticized lately. They indeed ignore the connections among the different layers, hence giving
low accuracy. To overcome these issues, we propose new community detection methods based on tailored
Louvain-like strategies that simultaneously handle the multiple layers. We consider the informative case,
where all layers show a community structure, and the noisy case, where some layers only add noise to the
system. We report experiments on both artificial and real-world networks showing the effectiveness of the
proposed strategies.

Key words. community detection, multi-layer networks, complex networks.

1 Introduction
Networks represented as graphs with nodes and edges have emerged as effective tools for modelling and
analysing complex systems of interacting entities. In fact, graphs arise naturally in many disciplines, such
as social networks [10], information networks [21], infrastructure networks [15], biological networks [8] and
many others.
One of the most relevant issues of graphs representing real systems is the identification of communities, or
clustering, i.e., the organization of vertices in clusters or modules. Although there is actually no consensus
on its definition, a community usually refers to a group of nodes that are compactly connected with each
other and sparsely connected with those nodes outside the group. Community detection in large networks
is potentially very useful [10, 8, 12, 18] and it is a very hard problem that is not yet satisfactorily solved,
despite the extensive studies in the literature.
While many community-detection and clustering algorithms have been developed over the recent years,
most of these are designed for standard single-layer graphs. On the other hand, having one single graph
is very often a oversimplifying assumption, which can lead to misleading models and results. Fortunato
conducted a survey on this topic [14].
∗
Department of Mathematics “Tullio Levi-Civita”, University of Padova, 35121 Padova, Italy (sara.venturini@math.unipd.it)
†
Department of Mathematics “Tullio Levi-Civita”, University of Padova, 35121 Padova, Italy (andrea.cristofari@unipd.it)
‡
Department of Mathematics “Tullio Levi-Civita”, University of Padova, 35121 Padova, Italy (rinaldi@math.unipd.it)
§
School of Mathematics, Gran Sasso Science Institute, Italy, (francesco.tudisco@gssi.it)
A. Cristofari, F. Rinaldi and F. Tudisco contributed equally to this work.

1
Recent advances in the study of networked systems have shown that the interconnected world is composed
of networks that are coupled to one another through different layers, where each layer represents one
of many possible types of interactions. Multi-layer graphs arise naturally in diverse applications such
as transportation networks [16], financial-asset markets [1], temporal dynamics [45, 46], semantic world
clustering [40], multivideo face analysis [6], mobile phone networks [20], social balance [7], citation analysis
[43], and many others.
Most of the current algorithms for community detection in multi-layer networks either reduce multi-layer
networks into a single-layer network or extend the algorithms for single-layer networks by using consensus
clustering. While these approaches perform well in the presence of well-clustered layers with communities
that are well in agreement, they are criticized for their low accuracy in more challenging settings such as
the case of noisy layers or poorly clustered ones, see e.g. [19, 42, 49, 54, 28].
In this paper, we propose two new strategies leading to multiple algorithms for community detection
in multi-layer networks. The algorithms are based on the popular Louvain heuristic method for single-
layer networks [3], which is a locally greedy modularity-increasing sampling process over the set of node
partitions, formed of two phases: a community evaluation phase and a node aggregation one. See Section 3
for details. The somewhat natural extension of this method to the multi-layer case, already studied in the
literature in e.g. [1], is to locally maximize a weighted average of the modularity of the layers during phase
1, instead of the modularity of a single layer. In Section 4.1, we suggest two variants that, in the selection
criteria of the algorithm during phase 1, take into account the variance of modularity on the layers, in
addition to the average. In Section 4.2, we present a more sophisticated filter-based type algorithm. Since
the problem under analysis can be viewed as a multiobjective optimization problem, where one tries to
optimize more than one objective function, we propose a technique to memorize and dynamically update a
list of several solutions, according to a suitably developed Pareto search. We make use of different criteria
for handling the list, that is, the modularity average over the layers and the convex combination of average
and variance, thus getting two different approaches. Notice that these methods with unit length of the
list correspond to the algorithms described before, respectively. Moreover, these algorithms coincide with
the Louvain method on single layer graphs. An important feature of both the proposed methods is that,
unlike many methods presented in the literature, they do not need to pre-assign the number of classes at
the beginning. This is particularly useful as we usually do not have information about the community
structure of a graph and one must make some a-priori (possibly unjustified) assumption about the number
of the clusters.
In order to verify the performance and the robustness of the various techniques, in our experimental set-up
we study multi-layer graphs in two settings: in the first one all the layers are informative, i.e. all layers
contain information about the community structure, whereas in the second setting at least one of the
layers is a noisy layer, which accounts for corrupted measurements. In Section 5.1, we provide extensive
experiments generating artificial multi-layer networks by a multi-layer version of the Stochastic Block
Model (SBM) [28], which show that our methods are competitive with those already proposed in the
literature. Moreover, in Section 5.2, we provide additional experiments on several real world graphs which
confirm our findings in the SBM.

2 Related work
In recent years, multi-view clustering has been a popular topic and many multi-view clustering algorithms
have been developed from different technical perspectives. In the following we attempt to review and
summarize some recent approaches.
Zeng et al. [50] proposed a pattern mining algorithm for finding closed quasi-cliques that appear on
multiple layers with a frequency above a given threshold. A cross-graph quasi-clique is defined as a set of

2
vertices belonging to a maximal quasi-clique that appears on all layers [35].
Tang et al. [44] and Dong et al. [11] proposed graph clustering algorithms for multi-layer graphs based
on matrix factorizations. The key point is to extract common factors from multiple graphs, based on
which general clustering methods can be applied. Tang et al. [44] approximates adjacency matrices while
Dong et al. [11] approximates graph Laplacian matrices. Liu et al. [25] proposed a non-negative matrix
factorization based multi-view clustering algorithm where the factors representing clustering structures
learnt from the multiple views are regularized toward a common consensus.
Another popular line of research tries to extend spectral clustering to multi-layer graphs. In general,
these algorithms aim to define a graph operator that contains all the information of the multi-layer graph
such that the eigenvectors corresponding to the smallest eigenvalues are informative about the clustering
structure. These methods usually rely on some sort of “mean operator”, for example the Laplacian of
the average adjacency matrix or the average Laplacian matrix [34]. Further examples are the work of
Zhou and Burges [55], which developed a multi-view spectral clustering via generalizing the usual single
view normalized cut to the multi-view case and tried to find a cut which is close to optimal on each layer,
and the algorithm designed by Chen and Hero [9], that performs convex aggregation of layers based on
signal-plus-noise models.
Zhan et al. [51], [52], [53] developed several multi-view graph learning approaches which are able to
merge multiple graphs into a unified graph with the desired number of connected components. Other
multi-view clustering approaches exploit the idea of maximizing clustering agreement. Zong et al. [56]
introduced Weighted Multi-View Spectral Clustering, where the largest canonical angle is used to measure
the difference between spectral clustering results of different views. Nie et al. [32] proposed a self-weighted
scheme for fusing multiple graphs with the importance of each view considered, called Procrustes Analysis
technique (AWP).
A common limitation to the proposed multi-view clustering methods is that they do not consider to deal
with possibly noisy or corrupted data, because they focus on the consistency of multiple layers and do not
consider the inconsistency. To address this issue, Xia et al. [48] proposed Robust Multi-view Spectral
Clustering, a Markov chain method that aims to learn an intrinsic transition matrix from multiple views
by restricting the transition matrix to be low-rank. This aspect has also been considered by Mercado et
al. [28] [27], where they propose a Laplacian operator obtained by merging the Laplacians from different
layers via a one-parameter family of nonlinear matrix power means. Recently, Liang at al. [24] proposed a
multi-view graph learning framework, which simultaneously models multi-view consistency and multi-view
inconsistency in a unified objective function.
Another line of work adopts Bayesian inference [47], in which certain hypotheses about connections between
nodes are made to find the best fit of a model to the graph through the optimization of a suitable likelihood
[36].
Bickel and Scheffer [2] extended the semi-supervised co-training approaches [4] to multi-view clustering.
The basic idea of co-training is to iterate over all views and optimize an objective function in the next
view using result obtained from the latter one. A co-training approach is proposed by Kumar and Daumé
[22], where the algorithm aims to find a consistent clustering that agrees across the views under the main
assumption that all the layers can be independently used for clustering. Under the same assumption,
Kumar et al. [23] proposed Co-regularized Spectral Clustering (CoReg), where they concentrated on this
approach under the notion of co-regularization, maximizing the agreement between different views.
Finally, some methods propose an extension of Newman’s modularity [30] and connected null models. For
instance, Bazzi et al [1] proposed a multi-layer modularity-maximization method, especially tailored to the
problem of identifying communities in multi-layer temporal networks, trying to find a trade-off between
static community structure within layers and larger values of persistence across layers. This approach
is related to our proposed strategy as it uses a cross-layer modularity function that is essentially based
on a weighted arithmetic mean of the modularities of the single layers. A related approach is proposed

3
by Pramanik et al [38] where a weighted linear combination of the modularities of each layer is used to
define a multi-layer modularity index. Their approach focuses on the case of two-layer networks with
both cross and within layer connections and aims at detecting cross-layer and within-layer communities
simultaneously. Mucha et al [29] generalize the Laplacian dynamics to derive null models for multi-layer
networks. They obtain a multi-layer generalization of modularity where the coupling between layers is
controlled using an additional parameter. Community detection on the multi-layer network then proceeds
using computational heuristics available for single layer graphs. Pizzuti and Socievole [37] develop a genetic
algorithm for community detection in multi-layer networks that makes use of a multiobjective optimization
formulation. In particular, the proposed algorithmic scheme exploits the concept of Pareto dominance
when creating the new populations at a given iteration, and returns a family of solutions that represent
different trade-offs between the objectives at the end of the optimization process (the best solution is
finally chosen using some tailored strategies).

3 The Louvain method for single layer graphs

In this work, we rely on the idea that a graph has a community structure if it is “different” enough from
a random graph. A random graph should indeed not have a community structure, since any two nodes
have the same probability to be adjacent. To this end, a null model is used as a term of comparison when
checking whether a graph shows a community structure or not. This is a core concept in the definition of
the modularity, a quality function that identifies a subgraph as a community if the number of edges inside
it exceeds the expected number of internal edges that the same subgraph would have in the null model.
The modularity function can be written as follows
1 X
Q= (Aij − Pij )δ(Ci , Cj ) (1)
2m
i,j

where the sum runs over all pairs of vertices, A is the adjacency matrix, m the total number of edges of
the graph (or the sum of all their weights in the case of weighted networks), Pij represents the expected
number of edges between vertices i and j in the null model, and Ci is the community of node i. The
function δ yields one if vertices i and j are in the same community (Ci =Cj ), zero otherwise.
Even though several variations of the modularity and the null model have been proposed in the literature,
the arguably most popular null model is the configuration model originally considered by Newman and
Girvan in [30], where edges are linked at random, under the constraint that the expected degree of each
vertex of the null model coincides with the actual degree of the corresponding node in the original graph
[31]. With this choice of P , the modularity function reads

1 X ki kj 
Q= Aij − δ(Ci , Cj ) (2)
2m 2m
i,j

where ki is the degree of the node i (the sum of all the edges incident to i ). With this model, communities
are found via the maximization of Q. As this is known to be a NP-hard combinatorial optimization task
[5], several algorithms have been proposed to approximately compute a modularity-based community
assignment.
The Louvain method, introduced by Blondel et al. in [3], is one of the most popular and most effective
methods for modularity maximization. This method is based on a greedy strategy that consists of two
phases: initially, each vertex is a community and, in the first phase, the algorithm computes the gain, in
terms of weighted modularity, obtained by including a node i in the community of every neighbour node j,
then selecting the one with the largest increase in modularity, as long as it is positive. Once the first level

4
partition is obtained, in the second phase of the method, communities are replaced by supervertices. Two
supervertices are connected if there is at least an edge between vertices of the corresponding communities,
with the edge weight between the two supervertices being the sum of the edge weights between the
corresponding communities. The whole procedure is repeated iteratively, until modularity cannot increase
and the algorithm stops.
The Louvain heuristic is very popular for its simplicity and efficiency. Part of the algorithm’s efficiency
results from the fact that the modularity can be calculated iteratively during the procedure. Only at the
very beginning of phase 1, the method calculates the modularity from scratch.
To calculate the modularity, the algorithm uses the following formula, that is equivalent to the equation
(2),
P !2
X |E(c)| ki
i∈N
Q= − (3)
m 2m
c∈C

where C =(C1 ,...,Cn ) is the clustering assignment, m is the sum of weights of all the links in the network,
|E(c)| is the sum of the weights of all the links between nodes in C and ki is the degree of node i.
During the loop, the algorithm calculates the modularity gain in a cheap way. The gain ∆Q1i , obtained
by moving a node i from its community c, can easily be computed via the formula
ki 2
P
·ki ki,in
∆Q1i = tot 2 − 2
− (4)
2m 2m m
P
where tot is the sum of weights of the links incident to the nodes in c, ki is the degree of node i, ki,in is
the sum of weights of the links from i to nodes in c.
Similarly, the following expression is used in order to evaluate the change of modularity ∆Q2i→j when an
isolated node i is moved into the community c of node j
P P
i,in ·ki
∆Q2i→j = − tot 2 (5)
m 2m
P
where i,in is the sum of weights of the links inside community c.
Thus, if a node i changes community, the algorithm calculates the modularity of the new partition just
by adding the gains obtained above to the initial modularity value, rather than calculating the function
Q from scratch. In particular, note that these formulas to calculate the modularity gain can be easily
extended to the weighted graph created in the second phase of the algorithm. Due to this fact, the Louvain
heuristic is extremely fast.

4 Louvain-like Methods for Multiple Layers

In this Section, we present our two methods for community detection in multi-layer networks: the Louvain
Expansion method and the Louvain Multiobjective method. They are both based on the above described
Louvain method for single-layer networks and they do not need a preassigned number of classes. We now
assume to have a so-called multiplex network: a multi-layer graph where all layers have the same nodes
and there are no edges between nodes of different layers. Each layer is an undirected graph and just a
single edge between nodes is allowed.

4.1 The Louvain Expansion Method

The idea behind the method is to define a suitable function F that combines the modularity score of the
individual layers and thus assesses the “quality” of a partition across multiple layers. In particular, we
consider two such functions F , which we describe next.

5
Algorithm 1 Louvain Expansion Method
G the initial multi-layer graph
Set Terminate = false
repeat
Put each node of G in its own community
while some nodes are moved do
for all node i of G do
place i in the community of one of its neighbors j, which maximizes the gain ∆Fi→j
end for
end while
if the new function value is higher than the initial then
G = the network between communities of G
else
Set Terminate=true
end if
until (Terminate)

Similarly to the standard Louvain method, starting with a partition where each node represents a
community, we proceed with a two-phase scheme. In the first phase, the algorithm picks one node at a
time, following a given initial node ordering. This ordering may affect the performance of the method as
will be further pointed out later on. For each node i, we compute the modularity gain ∆Q1i on each layer
obtained by removing i from its community Ci . For every neighbour j of node i that has not already been
considered, we also compute the modularity gain ∆Q2i→j on each layer obtained by including i in the
community Cj of node j, where we say that j is a neighbour of i if they are connected in at least one layer.
Using these quantities, we calculate the gain ∆Fi→j obtained from moving i from Ci to Cj , in a computational
efficient way. Then, we put i in the community Cj* which yields the largest gain ∆Fi→j* , provided the
latter is positive. The algorithm continues with the new partition (if any) and stops if the value of F
cannot be improved.
The second phase is the same as for the original Louvain method on single-layer graphs (see Section 3).
The whole procedure is repeated until no improvement can be found. The Louvain Expansion Method is
summarized in Algorithm 1.
As mentioned earlier, the initial ordering of the nodes can affect the final performance. In particular,
based on our computational experience, it seems that the ordering does not significantly affect the final
value of the cost-function F , while has more impact on the computational time. We would like to notice
that, even in the single layer case, this specific issue has not been properly investigated in the literature.
In our experiments, we hence choose the ordering depending on the specific network setting we deal with.
More specifically, we order the nodes according to the community size for the informative setting, while in
the noisy setting, which requires more computational effort, we list the nodes in random order.

4.1.1 Choices of F
It is reasonable to use a function F that includes the information of the multiple layers and whose variation
∆Fi→j can be iteratively calculated in an efficient way.
The most intuitive idea to extend the Louvain heuristic from the single-layer to the multi-layer case is to
consider as function F the average of the modularity values on the individual layers
Pk
Qs
MQ = s=1 (6)
k

6
where k is the number of layers and Qs is the modularity of layer s. The idea of considering a linear
(possibly weighted) combination of the modularity functions of the single layers is relatively natural and
has been considered for example in [1, 38].The gain of this function can be calculated easily as follow
Pk
s=1 ∆Qs
∆MQ = (7)
k
where k is the number of layers and ∆Qs is the gain on modularity of layer s. We refer to this method as
Louvain Expansion Average (EA).
While the average yields a reasonable overview of the graph community structure across the layers, it
is well known that in many situations this may lead to oversimplification [28, 27]. For example, in the
presence of noisy layers, linear averages over the multi-layer graph perform poorly [28]. To overcome this
issue, we consider two functions that embeds the sample variance of the modularity on the layers:

F- = (1 − γ)MQ − γVQ (8)

F+ = (1 − γ)MQ + γVQ (9)

where γ ∈ (0, 1) is a parameter, MQ is the average of modularity on the layers as in (7), and VQ is the
variance of modularity on the layers which we compute as
Pk
− MQ )2
s=1 (Qs
VQ = . (10)
k−1
The gain of both the functions can be calculated in an efficient way during the algorithm via the following
updating formulas
 2 
∆F- = (1 − γ)∆MQ − γ V∆Q + (Q − MQ )t (∆Q − ∆MQ ) (11)
k−1
 2 
∆F+ = (1 − γ)∆MQ + γ V∆Q + (Q − MQ )t (∆Q − ∆MQ ) (12)
k−1
where k is the number of layers, Q is a vector whose s-th component is the initial modularity of layer s,
MQ is the initial average of modularity on the layers, ∆Q is a vector whose s-th component is the gain in
modularity of layer s, ∆MQ is the gain of the modularity average on the layers (this coincides with M∆Q ,
the average of ∆Q, as shown in equation (7)), and V∆Q is the variance of ∆Q, calculated as follows:
Pk
s=1 (∆Qs − ∆MQ )2
V∆Q = . (13)
k−1
Note that if we iteratively update ∆Qs we then are able to iteratively update both (11) and (12).
We refer to the method that uses the function F- as Louvain Expansion Variance Minus (EVM) and to
the method that uses the function F+ as Louvain Expansion Variance Plus (EVP). These two strategies
enable us to take into account the variability of the modularity in the layers. We use EVM when dealing
with the informative case (no noise in the layers). The rationale behind this choice is to obtain a trade-off
between large modularity and sufficiently small variability in the layers (the larger the γ, the smaller the
variance in the final solution), which is what an ideal solution would look like if the community structure
of all layers is in agreement. When tackling noisy multi-layer networks, we instead use EVP. The idea in
this case is to handle the noise by allowing for solutions with a good level of variability in the layers and a
large enough modularity (a larger γ will get a larger variability in this case). All these methods coincide
with the original Louvain algorithm if applied to single-layer graphs.

7
4.2 The Louvain Multiobjective Method
Our second family of Louvain-like methods is the Louvain Multiobjective method for community detection
in multi-layer networks. This is a filter-type method that accounts for the multiobjective nature of the
problem under analysis, i.e., the existence of more than one objective to optimize. The main idea behind
this approach is to keep a list of partitions during the algorithm, each of them being preferable over the
others with respect to a specific criterion.
More formally, we consider a vector Q=(Qs )s=1,..,k where k is the number of layers, and we want to
maximize simultaneously all its entries. If there are no conflicts between the entries, the straightforward
optimal solution of this multiobjective problem is obtained by solving k optimization problems separately
(unfortunately this is usually not the case in real-world networks). In multiobjective optimization, there
is not a unique way to define optimality, since there is no a-priori total order for IRk , and each partial
order ends up giving a different definition. Let us hence provide a well known definition of optimality,
for a maximization problem, first proposed by Edgeworth in 1881 and later revised by Vilfredo Pareto in
1896 [33].

Definition 4.1. Given two vectors z1 and z2 ∈ IRk , z1 dominates z2 according to Pareto, and we write
z1 ≥ P z2 , if

z1i ≥ z2i for each index i = 1, .., k and

z1j > z2j for at least one index j = 1, .., k.

This binary relation induces a partial order over IRk . Thus we can give the definition of optimality
according to Pareto.

Definition 4.2. A vector z * ∈ IRk is Pareto optimal if there is no other vector z ∈ IRk such that z∗ ≤ P z.
Moreover, the Pareto front is the set of all Pareto optimal points.

Here, we still use a function F to assess the “quality” of a partition over several layers, but we also have a
filter L, that is, a list of partitions with community vectors such that no one is dominated by the others.
In the first phase, we initialize the method with the node-based partition (each node represents a community)
and we insert such a partition in the filter L. Then we start a loop, and analyse each node in a given order
(indexing of the nodes can again affect the performance of the algorithm). For every node i and for every
partition in the filter, we compute the gain ∆Q1i on each layer and, for every neighbour node j, also the
gain ∆Q2i→j on each layer, where ∆Q1i and ∆Q2i→j are defined as in the Louvain Expansion Method
described in Subsection 4.1.
Using these quantities, we calculate the gain ∆Fi→j in a computationally efficient way and, if this value
is positive, we include the new partition in the filter L. In particular, this happens if and only if the
modularity vector Q related to the partition is not dominated by any other modularity vector we have
with the partitions in L. Once the new partition is added to L we remove from L all the partitions whose
modularity vectors are dominated by the new one.
To prevent the filter from getting too large, we also cut down L to a predefined length h, if necessary, by
removing the worst partitions (with respect to the function F ). This operation is needed in order to ensure
the method computational efficiency.
The first phase then ends when the filter stops changing. At this point, the method selects from L the
best partition with respect to F and the second phase begins, which implements the same procedure as
the one described above for the Louvain method on single-layer graphs (see Section 4.1).
The whole procedure is repeated until no improvement is found. The Louvain Multiobjective Method is
summarized in Alg. 2.

8
Algorithm 2 Louvain Multiobjective Method
G the initial multi-layer graph
L initialized with node-based partition
Set Terminate = false
repeat
while L changes do
for all node i of G do
for all partition C in filter L do
place i in in every neighboring community which has a positive function F gain and put this
new partition in filter L
end for
end for
Cut the filter to length h
end while
Consider the partition in the filter L which maximizes the function F gain
if L has changed then
G = the network between communities of G
else
Set Terminate=true
end if
until (Terminate)

The order of nodes can affect the final output of this method as well, and, in our experiments, we make
the same choices reported in the previous section for the Louvain Expansion method.
We study three variants of the Louvain Multiobjective method, taking as function F : the modularity
average MQ described in (6), and the functions F- and F+ defined in (8) and (9). We refer to these
algorithms respectively as Louvain Multiobjective Average (MA), Louvain Multiobjective Variance Minus
(MVM), Louvain Multiobjective Variance Plus (MVP).
As we have seen in Section 4.1, the modularity and the function values can be calculated in an iterative
and easy way.
Note that this method coincides with the method described in Section 4.1 when the filter has a unit length.
Therefore, it is equal to the original Louvain algorithm if applied to single-layer graphs.

5 Experiments
We implemented the methods described in Section 4 using Matlab. The codes of all the methods are
available at the following link: https://github.com/ArasVentu/Louvain-like-Methods-for-Commun
ity-Detection-in-Multi-Layer-Networks.
We considered both synthetic and real-world networks, and performed extensive experiments to compare
the proposed methods against four state-of-the-art multi-layer community detection baselines, namely:
Co-Regularized Spectral Clustering (CoReg) (with parameter λ = 0.01) [23], multi-view clustering via
Adaptively Weighted Procrustes (AWP) [32], Multi-view Consensus Graph Clustering (MCGC) (with
parameter β = 0.6) [53], and The Power Mean Laplacian Multi-layer Method (PM) (with parameter
p = −10) [28]. Notably, all these algorithms require the user to specify the number of communities we
look for a-priori. This is a potential drawback in practice, as we usually do not have information about
the community structure of the graph and would have to make some (possibly unjustified) assumption

9
about the number of clusters.
One additional baseline method we compare with is the Louvain Expansion Average (EA), where we use
the Louvain scheme to optimize the modularity average (6), as discussed in Section 4.1. Using the Louvain
method for the (weighted) average of the modularity across the layers is a relatively established approach,
used for example in [1, 38].
We evaluated the performances of the various algorithms through two values: the accuracy (AC) (we report
the percentage of nodes placed in the right community) [53] and the Normalized Mutual Information
(NMI) [41].
For all experiment data, we considered two settings: the informative case, i.e., all the layers are informative,
and the noisy case, where some of the layers are just noise. For the informative case, we studied the
models Louvain Expansion Variance Minus (EVM) (Section 4.1), Louvain Multiobjective Average (MA)
and Louvain Multiobjective Variance Minus (MVM) (Section 4.2). For the noisy case, we studied the
models Louvain Expansion Variance Plus (EVP) (Section 4.1), Louvain Multiobjective Average (MA)
and Louvain Multiobjective Variance Plus (MVP) (Section 4.2).
In the definition of the function F- in equation (8) and F+ in equation (9), we set γ = {0.1, 0.3, 0.5, 0.7,
0.9} and we report the scores with the best parameter (i.e., the parameter achieving the highest NMI)
found on each dataset. As for the Louvain Multiobjective models, we consider two possible filter lengths:
h = {2, 3}, and we report the corresponding length right after the method name (e.g., MA2 stands for
the Louvain Multiobjective Average method with list length h = 2). In the noisy case, where both the
communities ordering and the noise in the layers are randomized, we report the average scores over 100
runs.

5.1 Synthetic Networks

We consider here networks generated via the (multi-layer) Stochastic Block Model (SBM), a generative
model for graphs with planted communities generated through the parameters p and q. These parameters
represent the edge probabilities: given nodes vi and vj , the probability of observing an edge between them
is p (resp. q), if vi and vj belong to the same (resp. different) cluster.
We set p > q in order to generate synthetic informative layers, while we simply let p = q for the noisy ones.
More precisely, we created networks with 4 communities of 125 nodes each and with k = 2, 3 layers, by
fixing p = 0.1 and varying the ratio p/q in the generation of the informative layers. In the noisy layers,
we fixed p = q = 0.1. For each value of the pair (p, q) we sample 10 random instances and we report
average scores. Results are shown in Figures 1 and 2, where we consider the following four settings: 1(a)
two informative layers; 1(b) three informative layers; 2(a) two informative layers and one noisy layer;
2(b) two informative layers and two noise ones. The multiobjective approaches M A/M V overall perform
best, with the variance-based one M V M or M V P being the best method overall. In general, the other
methods proposed here still outperform the baselines in almost all the parameter settings, reaching very
high accuracy and NMI even in the presence of two noisy layers (Fig. 2(b)). It is also interesting to notice
that, while the community detection problem becomes easier when the ratio p/q grows, our proposed
approaches still show performance advantages in that setting.

5.2 Real World Networks

We consider five real-world datasets frequently used for evaluation in multi-layer graph clustering: 3sources,
consisting of news articles covered by three different news sources [25]; BBCSports consisting of sports
news articles from [17]; Cora scientific publications with 7 classes [26]; Wikipedia, a dataset of Wikipedia
articles with ten classes [39] and the UCI multiple features digits dataset [13]. For each layer we build the
corresponding adjacency matrix from the k-nearest neighbour graph with k = 10 based on the Pearson

10
Table 1: Real-world datasets setting one. Average accuracy, NMI and performance ratio scores over 10
random initializations. All layers are informative. Best and second best values are highlighted with gray
boxes.

3sources BBCSport cora UCI Wikipedia Perf. Ratios

Acc NMI Acc NMI Acc NMI Acc NMI Acc NMI ρAcc ρN M I

EVM 0.876 0.789 0.833 0.798 0.617 0.537 0.882 0.921 0.548 0.520 0.955 0.967
MA2 0.858 0.749 0.899 0.825 0.407 0.434 0.753 0.862 0.525 0.521 0.863 0.912
MA3 0.876 0.789 0.596 0.731 0.425 0.428 0.876 0.910 0.558 0.546 0.844 0.917
MVM2 0.888 0.812 0.844 0.784 0.597 0.514 0.883 0.925 0.544 0.508 0.953 0.957
MVM3 0.888 0.812 0.915 0.851 0.603 0.502 0.883 0.925 0.530 0.504 0.965 0.967
EA 0.858 0.749 0.748 0.753 0.523 0.520 0.877 0.913 0.556 0.544 0.904 0.947
CoReg 0.651 0.658 0.858 0.617 0.530 0.380 0.958 0.911 0.522 0.445 0.889 0.805
AWP 0.686 0.662 0.616 0.722 0.534 0.293 0.869 0.891 0.462 0.332 0.805 0.753
MCGC 0.544 0.595 0.919 0.795 0.273 0.034 0.898 0.855 0.221 0.135 0.676 0.579
PM 0.734 0.707 0.778 0.690 0.551 0.456 0.876 0.879 0.569 0.560 0.896 0.896

Table 2: Real-world datasets setting two. Informative layers plus one noisy layer. Average accuracy,
NMI and performance ratio scores over 10 random initializations and 10 random edge probabilities
p ∈ {0.1, 0.3, 0.5} for the noisy layer. Best and second best values are highlighted with gray boxes.

3sources BBCSport cora UCI Wikipedia Perf. Ratios

Acc NMI Acc NMI Acc NMI Acc NMI Acc NMI ρAcc ρN M I

EVP 0.703 0.649 0.825 0.797 0.541 0.517 0.880 0.916 0.577 0.556 0.965 0.989
MA2 0.692 0.609 0.790 0.761 0.551 0.519 0.881 0.920 0.558 0.518 0.951 0.956
MA3 0.683 0.612 0.789 0.758 0.549 0.519 0.881 0.921 0.559 0.518 0.948 0.957
MVP2 0.717 0.668 0.828 0.797 0.543 0.518 0.881 0.920 0.578 0.555 0.971 0.996
MVP3 0.730 0.677 0.817 0.792 0.543 0.520 0.881 0.919 0.579 0.556 0.972 0.998
EA 0.678 0.607 0.777 0.754 0.555 0.521 0.881 0.920 0.561 0.520 0.947 0.955
CoReg 0.652 0.650 0.849 0.753 0.407 0.191 0.957 0.912 0.435 0.324 0.875 0.769
AWP 0.658 0.602 0.737 0.593 0.411 0.123 0.924 0.897 0.410 0.266 0.837 0.664
MCGC 0.546 0.585 0.812 0.694 0.303 0.005 0.804 0.816 0.201 0.107 0.687 0.565
PM 0.714 0.658 0.730 0.645 0.548 0.444 0.876 0.880 0.568 0.556 0.944 0.918

Table 3: Real-world datasets setting three. One aggregated informative layer plus one noisy layer. Average
accuracy, NMI and performance ratio scores over 10 random initializations and 10 random edge probabilities
p ∈ {0.1, 0.3, 0.5} for the noisy layer. Best and second best values are highlighted with gray boxes.

3sources BBCSport cora UCI Wikipedia Perf. Ratios

Acc NMI Acc NMI Acc NMI Acc NMI Acc NMI ρAcc ρN M I

EVP 0.655 0.575 0.886 0.799 0.550 0.432 0.869 0.903 0.556 0.526 0.965 0.954
MA2 0.348 0.217 0.664 0.555 0.541 0.418 0.853 0.890 0.431 0.361 0.775 0.721
MA3 0.332 0.207 0.659 0.550 0.537 0.416 0.858 0.893 0.427 0.359 0.768 0.716
MVP2 0.744 0.675 0.914 0.826 0.546 0.430 0.872 0.905 0.564 0.538 0.988 0.994
MVP3 0.754 0.689 0.914 0.828 0.544 0.431 0.873 0.906 0.566 0.540 0.991 1.000
EA 0.327 0.203 0.664 0.549 0.537 0.418 0.866 0.898 0.427 0.360 0.770 0.717
CoReg 0.566 0.436 0.608 0.338 0.435 0.190 0.761 0.642 0.446 0.305 0.768 0.551
AWP 0.549 0.416 0.644 0.376 0.444 0.179 0.750 0.622 0.439 0.292 0.770 0.540
MCGC 0.512 0.479 0.682 0.480 0.276 0.020 0.702 0.691 0.347 0.291 0.665 0.525
PM 0.512 0.363 0.729 0.658 0.569 0.414 0.834 0.828 0.457 0.360 0.848 0.772

11
 1 1 1 1
0.9 0.9
0.9 0.8 0.9 0.8
0.7

NMI

NMI
Acc

Acc
0.7
0.8 0.6 0.8
0.5 0.6
0.7 0.4 0.7 0.5
0.3 0.4
2 2.3 2.5 2.8 3 2 2.3 2.5 2.8 3 2 2.3 2.5 2.8 3 2 2.3 2.5 2.8 3
p/q p/q p/q p/q
EVM MA2 MA3 MVM2 MVM3 EVM MA2 MA3 MVM2 MVM3
EA CoReg AWP MCGC PM EA CoReg AWP MCGC PM

(a) (b)

Figure 1: The image is composed by two pairs (a) and (b) of line plots. They represent the average values
of accuracy and NMI over 10 random networks sampled from SBM with equally distributed informative
layers (2 layers (a) and 3 layers (b)), with four clusters, for p = 0.1 and p/q equal to 2, 2.3, 2.5, 2.8, 3.

1 1 1 1
0.9
0.9
0.9 0.8
NMI

NMI
Acc

0.9 0.8 Acc 0.7

0.8 0.6
0.7
0.5
0.8 0.6 0.7 0.4
2.5 2.8 3 3.3 3.5 2.5 2.8 3 3.3 3.5 2.5 2.8 3 3.3 3.5 2.5 2.8 3 3.3 3.5
p/q p/q p/q p/q
EVP MA2 MA3 MVP2 MVP3 EVP MA2 MA3 MVP2 MVP3
EA CoReg AWP MCGC PM EA CoReg AWP MCGC PM

(a) (b)

Figure 2: The image is composed by two pairs (a) and (b) of line plots. They represent the average values
of accuracy and NMI over 10 random networks sampled from SBM with both informative and noisy layers
(two informative and one noisy in (a) and two informative and two noisy in (b)). The informative layers
are equally distributed SBM graphs with four clusters, p = 0.1 and p/q equal to 2.5, 2.8, 3, 3.3, 3.5. The
noisy layers are SBM graphs with p = q = 0.1.

linear correlation between nodes, i.e. the higher the correlation the nearer the nodes are. In particular,
3sources is a dataset with 169 nodes per layer, 3 layers, and 6 communities of size 56, 21, 11, 18, 51, 12;
BBCSports is a dataset with 544 nodes per layer, 2 layers and 5 communities of size 62, 104, 193, 124, 61;
cora is a dataset with 2708 nodes per layer, 2 layers and 7 communities of size 298, 418, 818, 426, 217, 180,
351; UCI is a dataset with 2000 nodes per layer, 6 layers and 10 communities of 200 nodes each; Wikipedia
is a dataset with 693 nodes per layer, 2 layers and 10 communities of size 34, 88, 96, 85, 65, 58, 51, 41,
71, 104.
Given the ground truth community structure of the graphs, we analysed both the informative and the
noisy case. For the noisy case we study two settings. In the first one, we added a noisy layer to the
informative layers, whereas in the second setting we considered networks with 2 layers, where the first
one is the graph obtained aggregating all the layers and the other one is just noise. The noisy layers are
generated via uniform (Erdős–Rényi) random graphs with edge probability p ∈ {0.01, 0.03, 0.05}. For each
value of p, and each dataset, we generated 10 random instances and for each instance we run our methods
10 times with different random initial community orderings.
In Tables 1–3 we report average accuracy and NMI scores over all the samples and all the random
initializations. The best and second best values are highlighted with a gray box, with the best values

12
having a bold font. We further consider average performance ratio score values ρAcc and ρN M I , quantified
as follows: for a given measure Ma,d (where Ma,d is either Acc or NMI obtained by algorithm a over the
dataset d), the performance ratio is ra,d = Ma,d / max{Ma,d over all a}. The average performance ratios of
each algorithm ρAcc and ρN M I (for accuracy and NMI, respectively) are then obtained averaging ra,d over
all the datasets d. The closer is the average performance ratio of an algorithm to 1, the better will be the
overall performance of that algorithm.
In the last two columns of the Tables 1–3, we report the average performance ratios for all the tested
algorithms. We can see that in many cases the proposed methods overcome the baselines. In particular,
the methods that consider the variance in addiction to the average of the modularity across the layers
usually work better, with the multiobjective approaches M V M and M V P achieving best results in almost
all the cases. This is in agreement with the more sophisticated Pareto-based strategy and is consistent
with what observed in the SBM. Moreover, the last two columns highlight the robustness of the proposed
methods with respect to noise. In fact, in the nosy case (Tables 2 and 3) the proposed methods — in
particular the multiobjective ones — achieve very high performance ratios as well as overall high values of
accuracy and NMI.

6 Conclusions
In this paper, we presented new methods for community detection in multi-layer graphs that extend
the Louvain heuristic method in two different ways. The first method that we proposed, called Louvain
Expansion, combines the modularity average with the modularity variance, thus taking into account the
variability of the modularity in the different layers. The second algorithm, called Louvain Multiobjective,
is a more sophisticated filter-type method that maintains a list of suitably chosen modularity vectors
at each iteration and updates the corresponding community assignments according to a tailored Pareto
search. We considered different versions of these methods to better analyse two situations: the informative
case, where each layer shows some community structure, and the noisy case, where some layers present a
community structure and all the others are just noise. We provided extensive experiments on artificial
networks using the Stochastic Block Model. In addition, we tested our methods on five real datasets from
the literature. In both cases, we studied the informative and the noisy setting. The experimental results
demonstrate that the proposed methods are competitive with state-of-the-art baselines. In particular, the
multiobjective approach with the modularity variance shows best performance in almost all the cases.

References
[1] Marya Bazzi et al. “Community detection in temporal multilayer networks, with an application to
correlation networks”. In: Multiscale Modeling & Simulation 14.1 (2016), pp. 1–41.
[2] Steffen Bickel and Tobias Scheffer. “Multi-view clustering.” In: ICDM. Vol. 4. 2004. Citeseer. 2004,
pp. 19–26.
[3] Vincent D Blondel et al. “Fast unfolding of communities in large networks”. In: Journal of statistical
mechanics: theory and experiment 2008.10 (2008), P10008.
[4] Avrim Blum and Tom Mitchell. “Combining labeled and unlabeled data with co-training”. In:
Proceedings of the eleventh annual conference on Computational learning theory. 1998, pp. 92–100.
[5] Ulrik Brandes et al. “On modularity clustering”. In: IEEE transactions on knowledge and data
engineering 20.2 (2007), pp. 172–188.
[6] Xiaochun Cao et al. “Constrained multi-view video face clustering”. In: IEEE Transactions on Image
Processing 24.11 (2015), pp. 4381–4393.

13
 [7] Dorwin Cartwright and Frank Harary. “Structural balance: a generalization of Heider’s theory.” In:
Psychological review 63.5 (1956), p. 277.
[8] Jingchun Chen and Bo Yuan. “Detecting functional modules in the yeast protein–protein interaction
network”. In: Bioinformatics 22.18 (2006), pp. 2283–2290.
[9] Pin-Yu Chen and Alfred O Hero. “Multilayer spectral graph clustering via convex layer aggregation:
Theory and algorithms”. In: IEEE Transactions on Signal and Information Processing over Networks
3.3 (2017), pp. 553–567.
[10] James Samuel Coleman. An Introduction to Mathematical Sociology. Collier-Macmillan, London, UK,
1964.
[11] Xiaowen Dong et al. “Clustering with multi-layer graphs: A spectral perspective”. In: IEEE Trans-
actions on Signal Processing 60.11 (2012), pp. 5820–5831.
[12] Yon Dourisboure, Filippo Geraci, and Marco Pellegrini. “Extraction and classification of dense
communities in the web”. In: Proceedings of the 16th international conference on World Wide Web.
2007, pp. 461–470.
[13] Dheeru Dua and Casey Graff. UCI Machine Learning Repository – Multiple Features Data Set. 2017.
url: https://archive.ics.uci.edu/ml/datasets/Multiple+Features.
[14] Santo Fortunato. “Community detection in graphs”. In: Physics reports 486.3-5 (2010), pp. 75–174.
[15] Riccardo Gallotti and Marc Barthelemy. “The multilayer temporal network of public transport in
Great Britain”. In: Scientific data 2.1 (2015), pp. 1–8.
[16] Riccardo Gallotti and Marc Barthelemy. “The multilayer temporal network of public transport in
Great Britain”. In: Scientific data 2.1 (2015), pp. 1–8.
[17] Derek Greene and Pádraig Cunningham. “A matrix factorization approach for integrating multiple
data views”. In: Joint European conference on machine learning and knowledge discovery in databases.
Springer. 2009, pp. 423–438.
[18] Roger Guimera and Luıés A Nunes Amaral. “Cartography of complex networks: modules and universal
roles”. In: Journal of Statistical Mechanics: Theory and Experiment 2005.02 (2005), P02001.
[19] Jungeun Kim and Jae-Gil Lee. “Community detection in multi-layer graphs: A survey”. In: ACM
SIGMOD Record 44.3 (2015), pp. 37–48.
[20] Niko Kiukkonen et al. “Towards rich mobile phone datasets: Lausanne data collection campaign”.
In: Proc. ICPS, Berlin 68 (2010).
[21] Balachander Krishnamurthy and Jia Wang. “On network-aware clustering of web clients”. In:
Proceedings of the conference on Applications, Technologies, Architectures, and Protocols for Computer
Communication. 2000, pp. 97–110.
[22] Abhishek Kumar and Hal Daumé. “A co-training approach for multi-view spectral clustering”. In:
Proceedings of the 28th international conference on machine learning (ICML-11). 2011, pp. 393–400.
[23] Abhishek Kumar, Piyush Rai, and Hal Daume. “Co-regularized multi-view spectral clustering”. In:
Advances in neural information processing systems. 2011, pp. 1413–1421.
[24] Youwei Liang et al. “Multi-view Graph Learning by Joint Modeling of Consistency and Inconsistency”.
In: arXiv preprint arXiv:2008.10208 (2020).
[25] Jialu Liu et al. “Multi-view clustering via joint nonnegative matrix factorization”. In: Proceedings of
the 2013 SIAM International Conference on Data Mining. SIAM. 2013, pp. 252–260.

14
[26] Andrew Kachites McCallum et al. “Automating the construction of internet portals with machine
learning”. In: Information Retrieval 3.2 (2000), pp. 127–163.
[27] Pedro Mercado, Francesco Tudisco, and Matthias Hein. “Generalized matrix means for semi-
supervised learning with multilayer graphs”. In: Advances in Neural Information Processing Systems.
2019, pp. 14877–14886.
[28] Pedro Mercado et al. “The power mean Laplacian for multilayer graph clustering”. In: arXiv preprint
arXiv:1803.00491 (2018).
[29] Peter J Mucha et al. “Community structure in time-dependent, multiscale, and multiplex networks”.
In: science 328.5980 (2010), pp. 876–878.
[30] Mark EJ Newman. “Modularity and community structure in networks”. In: Proceedings of the
national academy of sciences 103.23 (2006), pp. 8577–8582.
[31] Mark EJ Newman and Michelle Girvan. “Finding and evaluating community structure in networks”.
In: Physical review E 69.2 (2004), p. 026113.
[32] Feiping Nie, Lai Tian, and Xuelong Li. “Multiview clustering via adaptively weighted procrustes”.
In: Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data
mining. 2018, pp. 2022–2030.
[33] Vilfredo Pareto. “Cours d’économie politique, Rouge”. In: Lausanne, Switzerland (1896).
[34] Subhadeep Paul and Yuguo Chen. “Spectral and matrix factorization methods for consistent
community detection in multi-layer networks”. In: The Annals of Statistics 48.1 (2020), pp. 230–250.
[35] Jian Pei, Daxin Jiang, and Aidong Zhang. “On mining cross-graph quasi-cliques”. In: Proceedings of
the eleventh ACM SIGKDD international conference on Knowledge discovery in data mining. 2005,
pp. 228–238.
[36] Tiago P Peixoto. “Bayesian stochastic blockmodeling”. In: Advances in network clustering and
blockmodeling (2019), pp. 289–332.
[37] Clara Pizzuti and Annalisa Socievole. “Many-objective optimization for community detection in
multi-layer networks”. In: 2017 IEEE Congress on Evolutionary Computation (CEC). IEEE. 2017,
pp. 411–418.
[38] Soumajit Pramanik et al. “Discovering community structure in multilayer networks”. In: 2017
IEEE International Conference on Data Science and Advanced Analytics (DSAA). IEEE. 2017,
pp. 611–620.
[39] Nikhil Rasiwasia et al. “A new approach to cross-modal multimedia retrieval”. In: Proceedings of the
18th ACM international conference on Multimedia. 2010, pp. 251–260.
[40] Joao Sedoc et al. “Semantic word clusters using signed spectral clustering”. In: Proceedings of the
55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).
2017, pp. 939–949.
[41] Alexander Strehl and Joydeep Ghosh. “Cluster ensembles—a knowledge reuse framework for com-
bining multiple partitions”. In: Journal of machine learning research 3.Dec (2002), pp. 583–617.
[42] Shiliang Sun. “A survey of multi-view machine learning”. In: Neural computing and applications
23.7-8 (2013), pp. 2031–2038.
[43] Wei Tang, Zhengdong Lu, and Inderjit S Dhillon. “Clustering with multiple graphs”. In: 2009 Ninth
IEEE International Conference on Data Mining. IEEE. 2009, pp. 1016–1021.

15
[44] Wei Tang, Zhengdong Lu, and Inderjit S Dhillon. “Clustering with multiple graphs”. In: 2009 Ninth
IEEE International Conference on Data Mining. IEEE. 2009, pp. 1016–1021.
[45] Dane Taylor, Rajmonda S Caceres, and Peter J Mucha. “Super-resolution community detection for
layer-aggregated multilayer networks”. In: Physical Review X 7.3 (2017), p. 031056.
[46] Dane Taylor et al. “Enhanced detectability of community structure in multilayer networks through
layer aggregation”. In: Physical review letters 116.22 (2016), p. 228301.
[47] Robert Winkler. Introduction to Bayesian Inference and Decision. Probabilistic Publishing, Gainesville,
2003.
[48] Rongkai Xia et al. “Robust multi-view spectral clustering via low-rank and sparse decomposition”.
In: Proceedings of the AAAI conference on artificial intelligence. Vol. 28. 1. 2014.
[49] Chang Xu, Dacheng Tao, and Chao Xu. “A survey on multi-view learning”. In: arXiv preprint
arXiv:1304.5634 (2013).
[50] Zhiping Zeng et al. “Coherent closed quasi-clique discovery from large dense graph databases”. In:
Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data
mining. 2006, pp. 797–802.
[51] Kun Zhan et al. “Graph learning for multiview clustering”. In: IEEE transactions on cybernetics
48.10 (2017), pp. 2887–2895.
[52] Kun Zhan et al. “Graph structure fusion for multiview clustering”. In: IEEE Transactions on
Knowledge and Data Engineering 31.10 (2018), pp. 1984–1993.
[53] Kun Zhan et al. “Multiview consensus graph clustering”. In: IEEE Transactions on Image Processing
28.3 (2018), pp. 1261–1270.
[54] Jing Zhao et al. “Multi-view learning overview: Recent progress and new challenges”. In: Information
Fusion 38 (2017), pp. 43–54.
[55] Dengyong Zhou and Christopher JC Burges. “Spectral clustering and transductive learning with
multiple views”. In: Proceedings of the 24th international conference on Machine learning. 2007,
pp. 1159–1166.
[56] Linlin Zong et al. “Weighted multi-view spectral clustering based on spectral perturbation”. In:
Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 32. 1. 2018.

16