Hidden Community Detection

on Two-Layer Stochastic Models: A

Theoretical Perspective

Jialu Bao1 , Kun He2(B) , Xiaodong Xin2 , Bart Selman3 , and John E. Hopcroft3
1
Department of Computer Science, University of Wisconsin-Madison,
Madison, WI 50706, USA
2
School of Computer Science and Technology, Huazhong University of Science
and Technology, Wuhan 430074, China
brooklet60@hust.edu.cn
3
Department of Computer Science, Cornell University, Ithaca, NY 14853, USA

Abstract. Hidden community is a new graph-theoretical concept

recently proposed by [3], in which the authors also propose a meta-
approach called HICODE (Hidden Community Detection) for detect-
ing hidden communities. HICODE is demonstrated through experiments
that it is able to uncover previously overshadowed weak layers and
uncover both weak and strong layers at a higher accuracy. However,
the authors provide no theoretical guarantee for the performance. In
this work, we focus on theoretical analysis of HICODE on synthetic two-
layer networks, where layers are independent to each other and each layer
is generated by stochastic block model. We bridge their gap through
two-layer stochastic block model networks in the following aspects: 1)
we show that partitions that locally optimize modularity correspond to
grounded layers, indicating modularity-optimizing algorithms can detect
strong layers; 2) we prove that when reducing found layers, HICODE
increases absolute modularities of all unreduced layers, showing its layer
reduction step makes weak layers more detectable. Our work builds a
solid theoretical base for HICODE, demonstrating that it is promising
in uncovering both weak and strong layers of communities in two-layer
networks.

Keywords: Hidden community · Multi-layer stochastic block model ·

Modularity optimization · Social network

1 Introduction
Community detection problem has occurred in a wide range of domains, from
social network analysis to biological protein-protein interactions, and numerous
algorithms have been proposed, based on the assumption that nodes in the same
community are more likely to connect with each other. While many real-world
J. Bao—Portion of the work was done while at Cornell University.
c Springer Nature Switzerland AG 2020
J. Chen et al. (Eds.): TAMC 2020, LNCS 12337, pp. 365–376, 2020.
https://doi.org/10.1007/978-3-030-59267-7_31
366 J. Bao et al.

social networks satisfy the assumption, their communities can overlap in interest-
ing ways: communities based on schools can overlap as students attend different
schools; connections of crime activities often hide behind innocuous social con-
nections; proteins serving multiple functions can belong to multiple function
communities. In any of these networks, communities can have more structures
than random overlappings. For example, communities based on schools may
be divided into primary school, middle school, high school, college and gradu-
ate school layers, where each layer are approximately disjoint. This observation
inspires us to model real world networks as having multiple layers.
To simulate real-world networks, researchers also build generative models
such as single-layer stochastic block model G(n, n1 , p, q) (p > q). It can be seen
as Erdős-Rényi model with communities—G(n, n1 , p, 1) has n nodes that belongs
to n1 disjoint blocks/communities (we use them interchangeably in the follow-
ing), and any node pair internal to a community has probability p to form an
edge, while any node pair across two communities have q probability to form an
edge. We propose a multi-layer stochastic block model G(n, n1 , p1 , ..., nL , pL ),
where each layer l consists of nl disjoint communities, and communities in dif-
ferent layers are independent to each other. Each layer l is associated with one
edge probability pl , determining the probability that a node pair internal to a
community in that layer forms an edge. In this ideal abstraction, we assume
that each node belongs to exactly one community in each layer, and an edge is
generated only through that process, i.e. all edges outgoing communities of one
layer are generated as internal edges in some other layers. Note that our model
is different to the multi-layer stochastic blockmodel proposed by Paul et al. [6],
where they have different types of edges, and each type of edges forms one layer
of the network.
He et al. [3,4] first introduce the concept of hidden communities, remarked
as a new graph-theoretical concept [7]. He et al. propose the Hidden Commu-
nity Detection (HICODE) algorithm for networks containing both strong and
hidden layers of communities, where each layer consists of a set of disjoint or
slightly overlapping communities. A hidden community is a community most of
whose nodes also belong to other stronger communities as measured by metrics
like modularity [2]. They showed through experiments that HICODE uncovers
grounded communities with higher accuracy and finds hidden communities in
the weak layers. However, they did not provide any theoretical support.
In this work, we provide solid theoretical analysis that demonstrates the
effectiveness of HICODE on two-layer stochastic models. One important step in
HICODE algorithm is to reduce the strength of one partition when the partition
is found to approximate one layer of communities in the network. Since commu-
nities in different layers unavoidably overlap, both internal edges and outgoing
edges of remaining layers have a chance to be reduced while reducing one layer.
It was unclear how the modularity of remaining layer would change. Through
rigorous analysis of three layer weakening methods they suggested, we prove
that using any one of RemoveEdge, ReduceEdge and ReduceWeight on one layer
increases the modularity of the grounded partition in the unreduced layer. Thus,
we provide evidence that HICODE’s layer reduction step makes weak layers more
detectable.
 Hidden Community Detection on Two-Layer Stochastic Models 367

In addition, through simulation, we show that on two-layer stochastic block

model networks, partitions with locally maximal modularity roughly correspond
to planted partitions given by grounded layers. As a result, modularity opti-
mizing community detection algorithms such as Louvain [1] can approximate
layers fairly accurately in a two-layer stochastic block model, even when layers
are almost equally strong and non-trivially overlapped. This indicates the previ-
ous proof’s assumption that one layer of communities is reduced exactly is rea-
sonable. We also illustrate how the modularity of randomly sampled partitions
change as HICODE iterates, and our plots show that not only absolute mod-
ularity but also relative modularity of unreduced layers increases as HICODE
reduces one found layer.

2 Preliminary
In this section, we first introduce metrics that measure community partition
quality. Then, we summarize important components in HICODE, the iterative
meta-approach we are going to analyze, and in particular, how it reduce layers
of detected communities during the iterations. Also, we define the multi-layer
stochastic block model formally, and the rationale why it is a reasonable abstrac-
tion of generative processes of real world networks.

2.1 Modularity Metric

In determining plausible underlying communities in a network, we rely on metrics
measuring quality of community partitions. Usually, nodes sharing common com-
munities are more likely to develop connections with each other, so in single-layer
networks, we expect that most edges are internal to one grounded community,
instead of outgoing edges whose two endpoints belong to two communities. It
thus gives rise to metrics measuring the similarity between an arbitrary partition
and the grounded partition based on the fraction of internal edges to outgoing
edges. One widely-used metric of this kind is “modularity” [2]. We define the
modularity of one community in multi-layer networks as follows:
Definition 1 (Modularity of a community). Given a graph G = (V, E)
with a total of e edges and multiple layers of communities, where each layer
of communities partitions all nodes in the graph, for a community i in layer
l, let eill denote i’s internal edges, and eilout denote the number of edges that
have exactly one endpoint in community i. Let dil be the total degree of nodes in
community i (dil = 2eill + eilout ). Then the modularity of community i in layer l
 i 2
ei d
is Qil = ell − 2ell .
Roughly, the higher fraction of internal edges a community has among all
edges, the higher its modularity in graph, indicating that members in that com-
munity are more closely connected.
When optimizing modularity, the algorithm concerns the modularity of a
partition instead of one community. The modularity of a partition is defined as
follows, which is consistent with the original definition of Girvan et al. [2]:
368 J. Bao et al.

Definition 2 (Modularity of a partition/layer). Given a network G =

(V, E) with multiple layers of communities, for any layer l, say l partitions
all the nodes into disjoint communities {1, . . . , N }, then the layer modularity
N
is Ql = i=1 Qil .
Whether in single-layer network or multi-layer ones, the ground truth community
partition is expected to have high modularity when compared to other possible
partitions.

2.2 HIdden COmmunity DEtection (HICODE) Algorithm

Informally, given a state-of-the-art community detection algorithm A for single
layer networks, HICODE(A) finds all layers in multi-layer networks through
careful alternations of detecting the strongest layer in the remaining graph using
A and reducing found layers on the network. Given a network G = (V, E),
He et al. [3] proposed three slightly different methods for reducing layers in
HICODE:
1. RemoveEdge: Given one layer l that partitions G, RemoveEdge removes
all internal edges of layer l from G.
2. ReduceEdge: Given one layer l that partitions G, ReduceEdge approximates
the background density q of edges contributed by all other layers, and then
removes 1 − q fraction of internal edges of layer l from network G. We will
detail the computation of q after introducing multi-layer stochastic block
model.
3. ReduceWeight: This is the counterpart of ReduceEdge on weighted graphs.
Given one layer l that partitions network G, ReduceWeight approximates
the background density q of edges contributed by all other layers, and then
reduces the weight of all internal edges to a q fraction of its original values.
For detailed description of HICODE, see [3].

2.3 Multi-layer Stochastic Block Model

Before defining the general multi-layer Stochastic Block Model (SBM), consider
the case where there is exactly two layers.
Definition 3 (Two-layer Stochastic Block Model). A synthetic network
G(n, n1 , p1 , n2 , p2 ) generated by two-layer stochastic block model has n nodes,
where n, n1 , n2 ∈ N + , n1 , n2 ≥ 3. For l = 1 or 2, layer l of G consists of nl
planted communities of size sl = nnl with internal edge probability pl ∈ (0, 1].
Communities in different layers are grouped independently, so they are expected
to intersect with each other by r = n1nn2 nodes.

Each community of layer l is expected to have pl · 12 sl1 internal edges1 . The

model represents an ideal scenario when there is no noise and all outgoing edges
1
For simplicity, we allow self-loops.
 Hidden Community Detection on Two-Layer Stochastic Models 369

of one layer are the result of them being internal edges of some other layers. We
will detail the expected number of outgoing edges and the size of the intersection
block of layers in Lemma 1 in the next section.
For example, in G(200, 4, 5, p1 , p2 ), layer 1 contains four communities C11 =
{1, 2, ..., 50}, C12 = {51, 52, ..., 100}, C13 = {101, 102, ..., 150}, C14 = {151, 152, ...,
200}, and layer 2 contains five communities C21 = {1, 6, ..., 196}, C22 = {2, 7, ...,
197}, C23 = {3, 8, ..., 198}, C24 = {4, 9, ..., 199}, C25 = {5, 10, ..., 200}. Each com-
munity is modeled as an Erdős-Rényi graph. Each C1i in layer 1 is expected
to have 0.5 · 502 p1 internal edges, and each C2i in layer 2 are expected to have
0.5 · 402 p1 internal edges.
Each community in layer 1 overlaps with each community in layer 2. Each
overlap consists of 20% of the nodes of layer 1 community and 25% of the nodes
of layer 2 community. Figure 1 (a) and (b) show the adjacency matrix when nodes
are ordered by [1, ..., n] for layer 1, and [1, 6, ..., 196, 2, 7, ..., 197, 5, 10, ..., 200] for
layer 2, respectively (Here we set p1 = 0.12, p2 = 0.10). Fig. 1 (c) and (d) show
an enlarged block for each layer. Edges in layer 1 are plotted in red, edges in
layer 2 are plotted in blue and the intersected edges are plotted in green.

(a) layer 1 (b) layer 2 (c) a L1 block (d) a L2 block

Fig. 1. The stochastic blocks in two layers. (Color figure online)

More generally, we can define a multi-layer stochastic block model.

Definition 4 (Multi-layer Stochastic Block Model). A multi-layer
stochastic block model G(n, n1 , p1 , ..., nL , pL ) generates a network with L lay-
ers, and each layer l has nl communities of size nnl with internal edge probability
pl . All layers are independent with each other.

2.4 Background Edge Probability for Multi-layer SBM

Given a layer, observed edge probability within its grounded communities would
be higher than its grounded edge generating probability, because other layers
could also generate edge internal to this layer. When we are interested in the
grounded edge generating probability of a layer, we can consider edges gener-
ated by all other layers as background noise. Since layers are independent to each
other, these background noise edges are uniformly distributed among communi-
ties of layer l, so we can expect background noise edge probability the same on
node pairs either internal to or across layer l communities. Thus, the observed
edge probability p of communities in a layer l equals p + q − p · q, where p is the
370 J. Bao et al.

grounded edge generating probability of layer l, q is the observed edge probabil-

ity across layer 1 communities. Thus, we can estimate the actual edge probability
−
by p = p1−q
q.

3 Theoretical Analysis on Two-Layer SBM

In this section, we show that on networks generated by two-layer stochastic block
model, weakening one layer would not decrease the quality of communities in any
other layer even when they considerably overlap with each other. We will prove
on two-layer stochastic block models that absolute modularity of unreduced layer
must increase after performing RemoveEdge, ReduceEdge, or ReduceWeight.
For simplicity, we make the assumption that the base algorithm can uncover a
layer exactly – every time it finds a layer to reduce, it does not make mistakes
on community membership. This is a strong assumption, but later on we will
justify why our result still holds if the base algorithm only approximates layers
and why the base algorithm can almost always find some approximate layers.
For each community in layer l, let sl denote the size of each community in
layer l, and ml denote the number of node pairs in the community. Since we
allow self-loops, ml = 12 s2l . Also, with the assumption that all communities in
one layer are equal sized, their expected numbers of internal (or outgoing) edges
are the same. Thus, we can use ell , elout to respectively denote the expected
number of internal, outgoing edges for each community i in layer l. Then, let
dl = 2ell + elout denote the expected total degree of any community in layer l.
Lemma 1. In the synthetic two-layer block model network G(n, n1 , n2 , p1 , p2 ),
for a given community i in layer 1, the expected number of its internal edges as
well as outgoing edges, and layer 1’s modularity are as follows:
 
1 1
e11 = 1 − m1 p 1 + m1 p12 , (1)
n2 n2
p2
e1out = s1 (n − s1 ), (2)
n2
1 e1out
Q1 = 1 − − , (3)
n1 d1
where p12 = p1 + p2 − p1 · p2 . Symmetrically, given a community i in layer 2,
the expected number of its internal edges as well as outgoing edges, and layer 2’s
modularity are as follows:
 
1 1
e22 = 1 − m2 p 2 + m2 p12 , (4)
n1 n1
p1
e2out = s2 (n − s2 ), (5)
n1
1 e2out
Q2 = 1 − − . (6)
n2 d2
We omitted the proof because of the space limit.
 Hidden Community Detection on Two-Layer Stochastic Models 371

Lemma 2. For layer l in a two-layer stochastic block model, if the layer weak-
ening method (e.g. RemoveEdge, ReduceEdge, ReduceWeight) reduces a bigger
percentage of outgoing edges than internal edges, i.e. the expected number of
e e
internal and outgoing edges after weakening, ell , elout , satisfy elout
lout
< ellll , then
the modularity of layer l increases after the weakening method.

We omitted the proof because of the space limit.

For a synthetic stochastic block model network G with set of layers L, let Sl
be the set of edges whose underlying node pairs are only internal to layer l ⊆ L,
let Sl1 l2 be the set of edges internal to both layers l1 , l2 ⊆ L. Concretely, in the
two-layer stochastic block model, L = {1, 2}. S1 is the set of edges only internal
to layer 1, S2 is the set of edges only internal to layer 2, and S12 is the set of
edges internal to both layer 1 and layer 2.

Lemma 3. In a two-layer stochastic blockmodel network G(n, n1 , n2 , p1 , p2 ),

before any weakening procedure.

|S12 | + |S1 | 2
e11 = , e1out = |S2 |,
n1 n1
|S12 | + |S2 | 2
e22 = , e2out = |S1 |.
n2 n2
We omitted the proof due to space limit.
Using the above three lemmas, we can prove the following theorems.

Theorem 1. For a two-layer stochastic blockmodel network G(n, n1 , n2 , p1 , p2 ),

the modularity of a layer increases if we apply RemoveEdge on communities in
the other layer.

Proof. If we remove all internal edges of communities in layer 1, both |S12 | and
|S1 | become 0, then the remaining internal edges of layer 2 is e22 = n12 (|S12 | +
|S2 |
|S2 |) = n2 > 0. There is no outgoing edge of layer 2, so e2out = 0. Thus,
e2out e22
e2out =0< e22 ,
and applying Lemma 2, we have that the modularity of layer 2
after RemoveEdge on layer 1 Q2 > Q2 .
Similarly, the modularity of layer 1 after RemoveEdge on layer 2, Q1 , is
greater than Q1 .

RemoveEdge not only guarantees to increase the absolute modularity of layer

2 but also guarantees that layer 2 would have higher modularity than any possible
partition of n nodes into n2 communities in the reduced network.
Theorem 2. For a two-layer stochastic blockmodel network G(n, n1 , n2 , p1 , p2 ),
If no layer 2 community contains more than half of the total edges inside it after
applying RemoveEdge on layer 1, then layer 2 has the highest modularity among
all possible partitions of n nodes into n2 communities.
372 J. Bao et al.

Proof. After applying RemoveEdge on layer 1, there are no outgoing edges of

any community in layer 2. It means that for any community i, ei2out = 0 and
di2 = 2ei22 . Thus, the modularity of layer 2 is:

   2
ei22 di2
Q2 = Qi2 = −
e 2e
i∈layer 2 i∈layer 2
  
4e · ei22 − (2ei22 )2 e · e22 − (e22 )2
= = n2 .
4e2 e2
i∈layer 2

For any one partition, we can transform layer 2 partition to it by moving a series
of nodes across communities. Every time we move one node from one community
i to another community j, both ei2out , ej2out will increase by 1, ei22 will decrease
by 2 while ej22 remains the same. Let ei i
2out , e22 denote corresponding values after
all movements. The following always holds no matter how many times we move:
 
2 (ei22 − ei
22 ) = ei
2out
i∈layer 2 i∈layer 2

Now Q2 , the modularity of the new partition after moving, is:

  2
ei di
Q2 = 22
− 2
e 2e
i∈layer 2
 4e · ei  (2ei i 2
22 + e2out )
= 22
− .
4e2 4e2
i∈layer 2 i∈layer 2

Let ei22 − ei

22 = Δi . Because of (a + b) ≥ a + b for any a, b ≥ 0, we have:
2 2 2

 4e · ei  (2ei )2 + (ei )2

Q2 ≤ 2
22
− 22 2out
4e 4e2
i∈layer 2 i∈layer 2
 i   i 2
4e · e22 − 4(ei 22 ) −
2
(e2out )
= 2
 4e  
4e · (ei22 − Δi ) − 4(ei22 − Δi )2 − (ei 2out )
2
= 2
 4e
   2
8 Δi e22 − 4e · Δi − (ei
i
2out ) − 4
2
Δi
= Q2 +
4e2
    2
Let T abbreviate 8 Δi ei22 −4e· Δi − (ei 2out ) −4
2
Δi , then Q2 = Q2 + 4eT2 .
When no layer 2 community contains more than half of the total edges after
applying RemoveEdge on layer 1, i.e., ei22 ≤ 2e ,
   
T =8 Δi ei22 − 4e · Δi − (ei
2out ) − 4
2
Δ2i
   
≤4e · Δi − 4e · Δi − (ei
2out ) − 4
2
Δ2i ≤ 0.
 Hidden Community Detection on Two-Layer Stochastic Models 373

Finally, we have Q2 ≤ Q2 + 4eT2 ≤ Q2 . Hence, layer 2 has the highest modular-
ity among all possible partitions of n nodes into n2 communities. In this way,
RemoveEdge makes the unreduced layer easier for the base algorithm to detect.
Theorem 3. For a two-layer stochastic blockmodel network G(n, n1 , n2 , p1 , p2 ),
the modularity of a layer increases if we apply ReduceEdge on all communities
in the other layer.
Proof. In ReduceEdge of layer 1, we keep edges in the given community with
probability q1 = 1−
1−
p
 is the observed edge probability within the
q , where p
detected community and q is the observed background noise.
ReduceEdge on layer 1 would only keep q1 fraction of edges in S12 and S1 ,
so after ReduceEdge,
1 1
e22 = (|S2 | + |S12 | · q1 ) > (|S2 | + |S12 |) · q1 = e22 · q1 ,
n2 n2
2
e2out = |S1 | · q1 = e2out · q1 .
n1
e e
Thus, e2out
2out
< e2222
, and Lemma 2 indicates that Q2 < Q2 . Similarly, for the
modularity of layer 1 after ReduceEdge on layer 1, Q1 > Q1 .
Theorem 4. For a synthetic two-layer block model network G(n, n1 , n2 , p1 , p2 ),
the modularity of a layer increases if we apply ReduceWeight on all communities
in the other layer.
Proof. According to [3], ReduceWeight on layer 1 multiplies the weight of edges
in layer 1 community by q1 = 1 − 1− p
q percent. In weighted network, the weight
1− 
sum of internal edges of a community i in layer 2 is e22 = 2 u,v∈i wuv · Auv
1

where wuv is the weight of edge (u, v). By construction, ReduceWeight on layer
1 reduces weight of all edges in S12 or S1 , but does not change weight of edges
in S2 . Thus,
1  1 
ei
22 = wuv · Auv · q1 + wuv · Auv
2 2
u,v∈i, (u,v)∈S12 u,v∈i, (u,v)∈S2
⎛ ⎞
1  1 
>⎝ wuv · Auv + wuv · Auv ⎠ · q1
2 2
u,v∈i, (u,v)∈S12 u,v∈i, (u,v)∈S2

= ·
ei22 q1
1 
ei2out = wuv Auv
2
u∈i,v ∈i
/
1 
ei
2out = wuv Auv · q1 = ei2out · q1
2
u∈i,v ∈i
/

e2out e22
Thus, <
e2out and combined with Lemma 2, this proves that Q2 > Q2 , the
e22 ,
modularity increases after ReduceWeight.
Similarly, the modularity of layer 1 after RemoveEdge on layer 1, Q1 > Q1 .
374 J. Bao et al.

The analysis shows that weakening one layer with any one of the meth-
ods (RemoveEdge, ReduceEdge, ReduceWeight) increases the modularity of the
other layer. These results follow naturally from Lemma 2, which is in some way
a stronger claim that the modularity of the remaining layer increases as long as
a larger percentage of outgoing edges is reduced than internal edges.

4 Simulation of Relative Modularity

To show whether reducing layers makes other layers more detectable when run-
ning HICODE, we simulate how grounded layers’ relative modularity changes
as the weakening method iterates on two-layer stochastic block models, and
compare the grounded layers’ modularity value with other partitions’ modu-
larity values. The number of possible partitions of n nodes is exponential, so it
would be computationally unrealistic just to enumerate them, let alone calculate
modularity for all of them. So we employ sampling of partitions. We calculate
modularity for all sampled partitions and plot them on a 2-dimensional plane
based on their similarities with the grounded layer 1 and layer 2, and show the
modularity values through the colormap with nearest interpolation.

4.1 Sampling Method

We sample 2000 partitions similar to layer 1 (or 2) by starting from layer 1 (or
2), and then exchange a pair of nodes or change the membership of one node for
k = 1, ..., 500 times. We also include 1200 partitions that mixed layer 1 and layer
2 by having k randomly selected nodes getting assigned to their communities in
layer 1 and the rest 200 − k nodes getting assigned to their communities in layer
2. As planted communities in different layers are independent, this sampling
method gives a wide range of partitions while being relatively fast. To measure
the similarity between two partitions, we adapt normalized mutual information
(NMI) [5] for overlapping communities (The definition of NMI is in Appendix
C.). Partitions of nodes are inherently high-dimensional. To place them on 2-
dimensional plane for the plotting purpose, we use its NMI similarity with layer
1 as the x-coordinate, and NMI similarity with layer 2 as the y-coordinate.
At each iteration, We use the modularity optimization based fast commu-
nity detection algorithm [1] as the base algorithm to uncover a single layer of
communities.

4.2 Simulation on ReduceEdge

Figure 2 presents the simulated results on a two-layer block model G(600, 15,
12, 0.1, 0.12) using ReduceEdge as the weakening method. In this network, layer
2 is the dominant layer (communities are bigger and denser) and layer 1 is the
hidden layer. The modularity of layer 2 is 0.546, while the modularity of layer
1 is 0.398. We plot the modularity of the estimated layer and other sampled
partitions at different iterations of HICODE. On each subfigure, the dark red
 Hidden Community Detection on Two-Layer Stochastic Models 375

Fig. 2. Simulation results of ReduceEdge on G(600, 15, 12, 0.1, 0.12). (Color figure
online)

cross sign denotes where the estimated layer projects on the 2-dimensional plane.
Simulations using RemoveEdge and ReduceWeight yield similar results.
1. Initially, two grounded layers here have similar modularity values, contribut-
ing to the two local peaks of modularity, one at the right-bottom and the
other at the left-top.
2. (a): At iteration t = 0:, the base algorithm finds an approximate layer 2,
whose NMI similarity with layer 2 is about 0.90.
3. (b): After reducing that partition, the modularity local peak at the left-top
sinks and the modularity peak at right-bottom rises, and the base algorithm
finds an approximate layer 1 whose NMI similarity with layer 1 is about 0.89.
ReduceEdge then reduces this approximated layer 1 and makes it easier to
approximate layer 2.
4. (c) and (d): At t = 1, the base algorithm finds an approximate layer 2 having
0.97 NMI similarity with layer 2, which is a significant improvement. As that
more accurate approximation of layer 2 is reduced, the base algorithm is
able to find a better approximation of layer 1 too. In our run, it finds an
approximation that has 0.96 NMI similarity with layer 1.
5. (e) and (f): As HICODE iterates, at t = 2, the base algorithm is able to
uncover an approximate layer 2 with 0.98 NMI similarity, and an approximate
layer 1 with 0.97 NMI similarity.

5 Conclusion
In this work, we provide a theoretical perspective on the hidden community
detection meta-approach HICODE, on multi-layer stochastic block models. We
376 J. Bao et al.

prove that in synthetic two-layer stochastic blockmodel networks, the modular-

ity of a layer will increase, after we apply a weakening method (RemoveEdge,
ReduceEdge, or ReduceWeight) on all communities in the other layer, which
boosts the detection of the current layer when the other layer is weakened. A sim-
ulation of relative modularity during iterations is also provided to illustrate on
how HICODE weakening method works during the iterations. Our work builds a
solid theoretical base for HICODE, demonstrating that it is promising in uncov-
ering both hidden and dominant layers of communities in two-layer stochastic
block model networks. In future work, we will generalize the theoretical analysis
to synthetic networks with more than two stochastic block model layers.

References
1. Blondel, V.D., Guillaume, J.L., Lambiotte, R., Lefebvre, E.: Fast unfolding of com-
munities in large networks. J. Stat. Mech. Theory Exp. 2008(10), P10008 (2008)
2. Girvan, M., Newman, M.E.: Community structure in social and biological networks.
Proc. Natl. Acad. Sci. 99(12), 3807–3870 (2015)
3. He, K., Li, Y., Soundarajan, S., Hopcroft, J.E.: Hidden community detection in
social networks. Inf. Sci. 425, 92–106 (2018)
4. He, K., Soundarajan, S., Cao, X., Hopcroft, J.E., Huang, M.: Revealing multiple
layers of hidden community structure in networks. CoRR abs/1501.05700 (2015)
5. McDaid, A.F., Greene, D., Hurley, N.: Normalized mutual information to evaluate
overlapping community finding algorithms. arXiv preprint arXiv:1110.2515 (2011)
6. Paul, S., Chen, Y.: Consistent community detection in multi-relational data through
restricted multi-layer stochastic blockmodel. Electron. J. Stat. 10(2), 3807–3870
(2016)
7. Teng, S.H., et al.: Scalable algorithms for data and network analysis. Found. Trends
Theor. Comput. Sci. 12(1–2), 1–274 (2016)