This article has been accepted for publication in IEEE Transactions on Neural Systems and Rehabilitation Engineering.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2023.3277509

GENERIC COLORIZED JOURNAL, VOL. XX, NO. XX, XXXX 2023 1

Dynamic Community Detection for Brain

Functional Networks during Music Listening with
Block Component Analysis
Yongjie Zhu, Jia Liu, and Fengyu Cong, Senior Member, IEEE

Abstract— The human brain can be described as a com- continuously naturalistic tasks.
plex network of functional connections between distinct
regions, referred to as the brain functional network. Re- Index Terms— dynamic community detection; brain con-
cent studies show that the functional network is a dy- nectivity; module detection; generative model; EEG; tensor
namic process and its community structure evolves with decomposition; block term decomposition.
time during continuous task performance. Consequently,
it is important for the understanding of the human brain
to develop dynamic community detection techniques for I. I NTRODUCTION
such time-varying functional networks. Here, we propose
a temporal clustering framework based on a set of network
generative models and surprisingly it can be linked to Block
Component Analysis to detect and track the latent commu-
T HE functional architecture of the human brain can be
characterized as a neuronal-synchronized network of
interconnected brain regions [1], [2]. Many studies of elec-
nity structure in dynamic functional networks. Specifically, trophysiological brain networks have provided new insights
the temporal dynamic networks are represented within into human behavior and cognition [3]–[5]. Early research
a unified three-way tensor framework for simultaneously focused on static functional connectivity (FC) patterns over
capturing multiple types of relationships between a set of
entities. The multi-linear rank-(Lr , Lr , 1) block term decom- time based on the stationary assumption. Recently, growing
position (BTD) is adopted to fit the network generative evidence has shown temporal dynamics of FC networks over
model to directly recover underlying community structures multiple time scales during continuous task performance and
with the specific evolution of time from the temporal net- resting states [6]–[8]. These network dynamics are critical to
works. We apply the proposed method to the study of the brain functions [9], [10] and dysfunctions [11]–[13]. Although
reorganization of the dynamic brain networks from elec-
troencephalography (EEG) data recorded during free music brain networks dynamically fluctuate over time, FC networks
listening. We derive several network structures (Lr commu- tend to be temporally clustered into a finite number of putative
nities in each component) with specific temporal patterns connectivity states, that is, distinct connectivity modules (com-
(described by BTD components) significantly modulated munities or subnetworks) that transiently form and dissolve
by musical features, involving subnetworks of frontopari- during continuous task performance [14]–[16]. Most of the
etal, default mode, and sensory-motor networks. The re-
sults show that the brain functional network structures research on dynamic FC states concentrates on the transition
are dynamically reorganized and the derived community among whole-brain network profiles only considering con-
structures are temporally modulated by the music features. nectivity edges [14], [17], [18]. However, few studies focus
The proposed generative modeling approach can be an on temporally switching in the topological organization of
effective tool for describing community structures in brain functional brain networks such as the modular or community
networks that go beyond static methods and detecting the
dynamic reconfiguration of modular connectivity elicited by structure.
Evidence from network neuroscience studies demonstrates
complex topological structures of both structural and func-
Manuscript received x xxx 2023; revised xx xxx 2023; accepted xx
xxx 2023. Date of publication xx xxx 2023; date of current version xx xxx tional brain networks [19], [20], where the brain networks can
2023. This work was partly supported by the National Natural Science be decomposed into clusters of densely interconnected nodes
Foundation of China (Grant No. 91748105), the National Foundation (referred to as modules or communities) that are relatively
in China (No. JCKY2019110B009 & 2020-JCJQ-JJ-252) and the Fun-
damental Research Funds for the Central Universities in the Dalian sparsely connected with nodes in other communities/modules.
University of Technology in China (DUT20LAB303). (Corresponding These topological communities typically correspond to clus-
authors: Yongjie Zhu, Fengyu Cong) ters of anatomically neighboring and/or functionally related
Y. Zhu is with the Department of Computer Science, University
of Helsinki, Helsinki, Finland, and with the Department of Neuro- brain regions which are involved in specialized functional
science and Biomedical Engineering, Aalto University, Espoo, Finland components [18], [20], [21]. Numerous community detection
(e-mail:[email protected]). algorithms have been developed for identifying underlying
J. Liu is with the Department of Biomedical Engineering, Faculty of
Engineering, Lund University, Lund, Sweden (e-mail:[email protected]). community structures in brain networks. The most widely
F. Cong is with the School of Biomedical Engineering, Faculty of used method in neuroimaging analysis is to apply modularity
Medicine, and the School of Artificial Intelligence, Faculty of Electronic maximization to the static brain networks, where nodes are
Information and Electrical Engineering, Dalian University of Technology,
Dalian, China, also with the Faculty of Information Technology, Univer- partitioned into non-overlapping and densely inter-connected
sity of Jyväskylä, Jyväskylä, Finland (e-mail:[email protected]). modules by maximizing an objective function of modularity

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Transactions on Neural Systems and Rehabilitation Engineering. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2023.3277509
2 GENERIC COLORIZED JOURNAL, VOL. XX, NO. XX, XXXX 2023

[22]. Most of them are only suited for analysis of static or tensor decomposition to generate the community [34], [36].
single-layer networks to define candidate communities at a Additionally, these approaches with tensor decomposition fail
fixed time [23], but characterizing time-evolving networks to provide a good generative model for the dynamic brain
with community structures has received less attention. networks; more precisely, the physical interpretations of the
The related studies for time-evolving networks primarily factors related to nodes and the temporal dimension are
focus on identifying clusters of a set of snapshots, i.e. network unclear.
connectivity patterns, that repeat themselves across time [14], In this paper, to overcome the limitations mentioned above,
[15]. For example, Ou and his colleagues introduced an we introduce a framework based on a latent network generative
approach based on statistical state modeling to identify the model and block term decomposition (BTD) [44], a variant
network states through hierarchical clustering followed by a of tensor decomposition [45]–[47], for detecting dynamic
Hidden Markov Model (HMM) [24]. In a similar manner, Ma community evolution in time-varying brain networks during
et al. identified the network states and their transitions through music listening. We first formulate a generative model to
independent vector analysis and Markov modeling [25]. Vidau- characterize community structure in time-varying brain net-
rre and colleagues have recently developed multiple methods works, quantified by envelope correlation of EEG recorded
based on HMM and applied them to neuroimaging data, during free music listening. Then, we show how to link
suggesting that functional networks transiently reorganize on the generative model to BTD and use it to learn the la-
the timescale of milliseconds [15], [16], [26], [27]. Under the tent community structures. Specifically, temporal concatenated
subspace modeling approaches, principal component analysis connectivity matrices are organized into a three-way tensor.
(PCA) [28] and independent component analysis [29] are used Then, BTD with rank-(Lr , Lr , 1) term is applied to extract
to extract the FC patterns, where it is assumed that brain the underlying community structures with a specific temporal
networks are composed of eigenconnectivities or independent mode. Different from previous tensor decomposition such
components. An alternative popular approach is based on k- as CANDECOMP/PARAFAC (CP) model for brain network
means clustering of dynamic functional connectivity networks analysis, the factors with rank Lr are able to characterize
(dFCNs) across time to identify the FC-states during rest, Lr communities after the multi-linear rank-(Lr , Lr , 1) BTD,
where it is assumed that a finite number of FC patterns recur which can discover the topological structures of brain net-
across time [14]. Although these methods were beneficial to works. Actually, the CP model can be considered as rank-
summarize the overall dynamic brain activity, they failed to un- (1, 1, 1) decomposition and it is unable to reveal the com-
cover the topological properties of the whole-brain networks. munity structures so further analysis is needed for the loading
Although dynamic community detection technique [30] has factors related to the node. After BTD, time series of five long-
recently emerged as a powerful tool for tracking the topo- term acoustic features were extracted from the audio stimuli
logical reconfiguration of brain networks [18], [31]–[33], it is by music information retrieval techniques used in previous
still not straightforward for module detection for time-varying studies [9], [48]. Finally, we analyzed the correlation between
networks within or across multiple subjects. We thus con- temporal factors and the musical feature time series to identify
sider the tensor decomposition (or tensor component analysis) underlying community structures of brain networks modulated
based methods for such dynamic community detection [18], by musical features.
[31], [34]–[37] since the tensor decomposition enables multi- The main contributions of this work are three-fold. First,
timescale dimensionality reduction both within and across we propose a generative model to characterize the temporal
temporal evolution for multiple subjects in a purely data-driven community evolution for dynamic brain functional networks.
method. Tensor decomposition has recently been regarded as Second, we show the proposed generative model can be
an extension of PCA for dynamic brain network analysis fitted with multilinear rank-(Lr , Lr , 1) BTD to learn the
across subjects, where time-frequency vectorized adjacency latent community structures in the time-varying brain networks
matrices were formed into a tensor and decomposed into com- without further analysis of the resulting BTD components.
ponents characterizing brain network patterns with spectral- Third, the proposed framework is then used for EEG networks
temporal features [9], [38]–[40] or temporal features [41], during naturalistic music listening to identify music-modulated
[42]. However, in this case, the topological organization of community structures, which demonstrate its effectiveness in
brain networks is unable to be directly revealed in the resulting time-varying modular detection for brain networks.
components. For such community detection in temporal brain
II. M ATERIALS AND M ETHODS
networks, tensor-based approaches typically model a network
as a three-way tensor and apply low-rank tensor decomposition A. Notation
to extract latent components [42], [43]. Each component is In this paper, scalars are denoted by lowercase letters
made up of three factors named “loading factors”. Two of the (a, b, · · · ), vectors are denoted by boldface lowercase letters,
factors relate to nodes and are used to generate a community such as (a, b, · · · ), matrices are written in boldface uppercase
with clustering or binary classification [43]. The other loading letters (A, B, · · · ), and high order tensors by boldface calli-
factor contains temporal information for tracking the temporal graphic letters (A, B, · · · ). Operator ◦ represents outer product
evolution [42]. Despite that these tensor-based methods have of vectors, ⊗ denotes the partitionwise Kronecker product, ⊙
achieved success in dynamic community detection for brain represents Khatri–Rao product and ⊙c denotes columnwise
network analysis, further analysis, such as k-means clustering Khatri–Rao product [49]. The superscripts ·T and ·† indicate
or classification of node loading factors, is required after the transpose and Moore–Penrose pseudoinverse, respectively.

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Transactions on Neural Systems and Rehabilitation Engineering. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2023.3277509
ZHU et al.: DYNAMIC COMMUNITY DETECTION WITH BLOCK COMPONENT ANALYSIS 3

B. Data Description and Preprocessing the number of samples.

EEG data from 14 right-handed adults between the ages
of 20 and 46 were used in the current study. No participant C. Dynamic Functional Connectivity Network
reported a history of hearing loss or neurological disease and Construction
none of them had music expertise. This research was approved
by the local ethics committee and has no conflicts of interest. We attempt to obtain an all-to-all whole-brain FC network
We presented subjects with a piece of music, which was played by computing connectivity between all pairs of DKA regions.
via audio headphones. The used music was a 512-second In M/EEG, a significant confound of electroencephalography
long musical segment of modern tango, which had a suitable source connectivity is that the ill-posed inverse problem and
duration for the experimental setting due to its high range inaccuracies in the forward solution lead to a degree of
of fluctuation in several musical features [9], [50]. EEG data spatial ambiguity and mislocalization of source [29], [56].
were collected at a sampling rate of 2048 Hz with BioSemi In other words, two source-level time signals (e.g. from two
electrode caps of 64 channels when participants were naturally brain regions) might be significantly correlated, merely due
listening to the continuous musical segment. to ‘signal leakage’ [17]. The obtained connectivity between
In this paper, we studied five well-known long-term acoustic spatially separate brain areas might be inaccurate without
features consisting of tonal and rhythmic features, which were careful control. To solve this issue, we performed the or-
computed by using a frame-by-frame analysis technique [48], thogonalization of source-reconstructed signals, a widely used
[50]. we set the length of each frame as three seconds and technique for leakage reduction [29]. Following signal leakage
the overlap between adjacent frames as two seconds. Thus, a correction, the Hilbert transformation was applied to extract
time course with 510 samples was created for each musical the amplitude envelopes of the time courses. The dynamic
feature with a sampling rate of 1 Hz. The five acoustic features FC networks were constructed for each subject by calculating
include two tonal musical features, Mode and Key Clarity, and the Pearson correlation, X ∈ Rn×n , between different n = 68
three rhythmic features, consisting of Fluctuation Centroid, DKA regions using a sliding window approach [14]. The rect-
Fluctuation Entropy, and Pulse Clarity. angular window length was set as 3 seconds and the overlap
was 2 seconds between two adjacent windows, resulting in
In preprocessing steps, EEG data were re-referenced by
a sampling rate of 1 Hz in the temporal dimension. This
common average electrodes and were visually inspected to
sampling rate was in line with the musical feature time series.
reject typical artifacts. We interpolated bad channels with a
mean value of their spherical adjacent channels. We used a 50 For each subject, a sequence of functional brain networks
Hz notch filter to remove powerline interference. High-pass G = {Gt (V, Et )|t = 1, · · · , τw } is constructed, where Gt
and low-pass filters with 2 Hz and 35 Hz cutoffs were then is the network snapshot at time t, V = {vi }ni=1 is a set of
applied since our previous investigation of the frequency range n nodes and Et is the set of edges at time step t. Here,
revealed that no useful information was observed in higher the nodes are the 68 DKA regions (n = 68), and the edges
frequencies [48], [51]. Finally, we down-sampled the EEG Et are constructed with Xt , which is the adjacency matrix
data to 256 Hz. Independent component analysis (ICA) was representing the network at time step t. τw = 510 is the
applied to individual EEG data to remove EOG artifacts (e.g. total number of time windows. The adjacency matrices were
eye blinks) [52]. temporally concatenated across subjects, resulting in a group-
level functional brain network G with τ = τw × np , where np
The schematic diagram of subsequent data processing is
is the number of subjects.
shown in Figure 1. Following data preprocessing, we estimated
the forward model and inverse model using a MATLAB-based
toolbox Brainstorm [53]. The symmetric boundary element
D. Generative Model for Dynamic Functional Networks
method (BEM) was applied to compute the forward model
with a default MNI MRI template (Colin 27). To solve the To allow module detection, we here extend the temporal
inverse model, we used weighted minimum-norm estimate clustering model under the CP framework in [34] to a gener-
(wMNE) [54]. The reconstructed cortical surface was deci- ative model under the BTD framework. We model a dynamic
mated to 4098 evenly distributed vertices per hemisphere with network functional brain networks G = {Gt (V, Et )|t =
4.9 mm spacing. Depth-weighted L2-minimum-norm estimate 1, · · · , τ } as a mixture of R generative models{S r }R r=1 . S
r

was computed for all current dipoles with a loose orientation contains the same set of nodes V , for r = 1, · · · , R, with
of 0.2. The inverse solution was noise-normalized. Then, the Lr communities (or subnetworks/modules) {Clrr }L r
lr =1 . We
cortical surface was parcellated into 68 anatomical regions assume that the probability that a node i ∈ V belongs to a
based on the Desikan-Killiany Atlas (DKA) [55]. For each community Clrr in S r follows a Bernoulli distribution, denoted
parcel, we performed a principal component analysis to extract as P (i ∈ Clrr ) = arilr . At time step t, the lth community in
orthogonal components that describe the activity, ordered by the rth generative model S r generates the connection between
amount of variance explained. We selected the first principal two nodes i, j with probability arilr arjlr λrt , where λrt is defined
component as a representation of the parcel’s time course of as an connection-generating rate or strength. λrt changes
activity. Thus, for each subject, a source-level data matrix throughout time and can be modeled as a time series, which
P was created with dimension n × nt , where n = 68 here represents the temporal evolution of the communities.
represents the number of anatomical regions and nt represents Consequently, the rth generative model S r generates the edge

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Transactions on Neural Systems and Rehabilitation Engineering. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2023.3277509
4 GENERIC COLORIZED JOURNAL, VOL. XX, NO. XX, XXXX 2023

Fig. 1. Analysis pipeline. EEG data were recorded during continuous music listening and then source-localized with wMNE. Source-localized
data were parcellated into 68 ROIs based on an anatomical brain template. After signal leakage correction, the Hilbert transformation was applied to
extract the amplitude envelopes of the ROIs’ time courses. An adjacent matrix was thus obtained by computing the correlation between the envelope
of separate regions for each time window. Then a three-way tensor was formed including two node modes and a temporal mode. Nonnegative BTD
decomposition was applied to the temporally concatenated tensor across subjects. The node factor matrix of extracted components with rank-Lr
is able to characterize the topological structures of the latent network patterns, which encodes Lr communities or node clusters, and the temporal
courses represent the time evolution of the modular patterns. On the other hand, musical features were extracted using acoustic feature extraction.
The temporal courses of decomposed components and musical feature time series were analyzed to examine the modulated brain networks.

with the sum across Lr communities at time step t: We suppose that the number of the generative components
R and the community number Lr , r = 1, · · · , R in each
Lr
X component are given here merely for simplicity of exposition.
(aril1 arjl1 +· · ·+arilr arjlr +· · ·+ariLr arjLr )λrt = (arilr arjlr )λrt
lr =1
(1) E. Learning Dynamic Community Structures with
which could be considered as the expected number of connec- Rank-(L, L, 1) BTD Model
tions generated between i and j in S r at time t. PTo estimate
PLr the community structure in Eq. P(3), we rewrite
R r r r R r rT r
Align and compact adjacency matrices, Xt , temporally into r=1 ( (a a
lr =1 ilr jlr )λ t ) in Eq. (3) to r=1 (ai aj λt )
r Lr
the third mode of a third-order tensor, namely X::t = Xt , with vector format, where ai ∈ R . The optimization
t = 1, · · · , τ . Element Xijt can be interpreted as the number problem in Eq. (3) can further be rewritten as,
of connections (the strength of connections) observed between R
X
node i and j at time step t. Thus, the model approximates Xijt min ∥X − (Ar · ATr ) ◦ cr ∥2F
Ar ,cr
by summing the Eq. (1) across R generative models as follow: r=1 (4)
R X Lr
s.t. 0 ≤ Ar ≤ 1; for r = 1, · · · , R
cr ≥ 0
X
Xijt ≈ ( (arilr arjlr )λrt ) (2)
r=1 lr =1 where Ar := [ar1 · · · arn ] ∈ Rn×Lr , cr := [λr1 · · · λrτ ]T ∈ Rτ .
In other words, the dynamic community detection problem Surprisingly, the optimization problem in Eq. (4) can be solved
is here to find R network generative models S r (or latent by using the BTD framework with rank-(Lr , Lr , 1) [44]. The
source components), their temporal evolution (connection- multilinear rank-(Lr , Lr , 1) terms decomposition factorizes a
generating) λrt , and the probability that node i belongs to three-way tensor into a sum of R low multilinear rank terms,
the one of communities in the generative source S r , arilr , each of which can be written as the outer product of a rank Lr
for r = 1, · · · , R; t = 1, · · · , τ ; lr = 1, · · · , Lr ; and matrix and a vector as shown in Eq. (4). Loading matrix Ar
i = 1, · · · , n. We can thus formulate the objective function with rank-Lr is able to characterize the modular structures of
as follow: the latent network pattern, which encodes Lr communities or
node clusters since the elements represent the probability that
R X Lr
X X X the n nodes belong to which communities. The node cluster
min ∥Xijt − ( (arilr arjlr )λrt )∥2F of each community can be obtained by the largest entry in the
i,j∈V t r=1 lr =1
(3) corresponding row of Ar . That is, if k = arg maxj {(Ar )ij },
r
s.t. 0≤ ailr ≤ 1; for i = 1, · · · , n then node vi belongs to k th node cluster. cr characterizes the
λrt ≥ 0 temporal evolution of the community structures (see Fig. 1).

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Transactions on Neural Systems and Rehabilitation Engineering. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2023.3277509
ZHU et al.: DYNAMIC COMMUNITY DETECTION WITH BLOCK COMPONENT ANALYSIS 5

Like CP-based tensor decomposition, there are many ap- number of latent components and rank. The data fit can be
proximate algorithms for rank-(Lr , Lr , 1) BTD decomposi- obtained as
tion, such as multiplicative updating (MU) method, alternating PR
∥X − r=1 (Ar · ATr ) ◦ cr ∥2F
least squares (ALS) and hierarchical alternating least squares F it(R, Lr ) = 1 − (6)
(HALS) [57]. Here, we use the structured data fusion ALS ∥X ∥2F
implementation by the Tensorlab [58], which is so far the most We can fix one of them, R or Lr to examine the changes
widely used computation scheme for the BTD model. The in the data fitting. Unlike PCA, the estimation of BTD may
ALS algorithm applies a gradient descent method to solve the have local minima (suboptimal solution), and not guarantee
minimization problem in Eq. (4) iteratively. At each iteration, that optimization routines will converge to the global optimal
one of the factor matrices is updated while other factor solution. Thus, we run the ALS optimization procedure at
matrices are fixed. We define A := [A1 · · · AR ] ∈ Rn×R̂ , each component number R or each rank value Lr 20 times
PR
C := [c1 · · · cR ] ∈ Rτ ×R̂ , where R̂ = r=1 Lr . For brief from random initial conditions. Generally, the candidate model
illustration, consider estimating A, fixing C, resulting in the order Re and rank L fr can be thought of as the appropriate
following update rule: selection when the data fitting no longer increases as the
R number increases.
X
A ← arg min ∥X − (Ar · ATr ) ◦ cr ∥2F , for fixed C
0≤A≤1
r=1 G. Temporal Modulation of the Community Structure by
(5)
Musical Features
It can be estimated as a linear least-squares problem and
has a closed-form solution. The update procedure for A To examine how musical features temporally modulate the
and C is summarized in Algorithm 1. The detailed solution topological (module) structures of brain functional networks,
of rank-(Lr , Lr , 1) terms decomposition using ALS and its we here adopt temporal modulation analysis for each musical
convergence analysis can be found in [44], [49]. feature, time courses of modular structures (components), and
subject. Previous studies have shown that the topological
Algorithm 1: ALS organization of functional networks temporally evolves to
support ongoing cognitive function [14], [29]. We here attempt
Input: ALS(X ∈ Rn×n×τ , R, L1 , L2 , · · · , LR )
to perform a correlation analysis between the temporal courses
Output: A ∈ Rn×R̂ , C ∈ Rτ ×R̂ of modular patterns and musical time series, by assessing
1 begin
2 initiazation for A, C the statistical significance of temporal correlations based on
3 repeat a surrogate permutation procedure [9], [29]. We obtain R
4 Update A: BTD components with two loading factors, characterizing
5 Ae = [(A ⊙ C)† · (A ⊙ C) · AT ]T + the temporal evolution (represented by cr ) and topological
6 for r = 1 to R do structures of brain networks (represented by Ar with rank Lr ).
7 A
e r = QR QR-factorization
The temporal factor matrix C (C ∈ Rτ ×R ) is first reshaped
8 Ar ← Q
as a three-way tensor C (C ∈ Rτw ×np ×R ). which consists of
9 end
10 A = [A1 · · · AR ] an individual time course for each BTD component. For each
11 Update C: BTD component and each subject, we calculate the correlation
12 T = (A1 ⊙c A1 )1L1 · · · (AR ⊙c AR )1LR coefficients between each musical feature time series and time
13 C ← [T† · T · CT ]T+ courses as the modulation scores. We then evaluate which BTD
14 until convergence; component is significantly modulated by examining whether
15 end;
its modulation score is significantly different from the scores
of surrogate data. We generate the surrogate data with a phase-
randomization procedure [60], which randomizes the intrinsic
phase and retains the properties of the temporal course in
F. Model Order Selection the spectral domain. The phase-randomization procedure is
When estimating the multi-linear rank-(Lr , Lr , 1) BTD repeated 5000 times for each BTD component. We compute
model, a natural question follows: how to select R and Lr the correlation coefficients between musical feature time series
from the experimental data? So far there is unfortunately still and phase-randomized time courses to generate a distribution
no gold standard method for model order selection of the of modulation scores from surrogate data. The 95th percentile
BTD model in the literature [59]. A common practice is to (p = 0.05) of surrogate modulation scores are chosen as
determine R and Lr that result in rational decomposition the threshold (the control modulation score for comparisons)
results according to the data fitting values and the prior for each subject. This significant level was corrected based
knowledge of data features. Here, we use the model fitting on Bonferroni correction for multiple comparisons across
method, based on the measurement of the data fitting, as a the multiple components i.e., pcorrect = 0.05/R. For each
reference to choose the model order. Data fitting is computed component, we finally use two-tailed t-tests for the modulation
based on model reconstruction error and the explained variance score of each musical feature to determine which component
of data. Let component number R ∈ [1, R] and the rank (brain network pattern) is modulated significantly differently
Lr ∈ [1, L], where R and L are the empirically maximal at pcorrect = 0.05/R level from the defined threshold.

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Transactions on Neural Systems and Rehabilitation Engineering. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2023.3277509
6 GENERIC COLORIZED JOURNAL, VOL. XX, NO. XX, XXXX 2023

Fig. 2. Results of simulation data. Left: the modular structures and temporal profiles of three synthetic brain network patterns. Right: the
corresponding modular structures and temporal profiles of reconstructed brain patterns. I, II, and III represent the three components respectively.
Note that the scale of the amplitude is different between simulated and reconstructed temporal profiles due to the scaling indeterminacy in tensor
decomposition.

III. R ESULTS showed the case under the signal-to-noise ratio (SNR) of 10dB.
A. Simulation Results One can observe that the three latent brain network patterns
with distinct topological structures and temporal modes were
We first validated the proposed approach with simulation
successfully extracted using multilinear rank-(Lr , Lr , 1) BTD.
data, which provided the instruction to study the performance
of the methodology. The performance of the Pearson corre-
lation of the envelope with signal leakage reduction, as a B. Rsults from EEG data recorded during music listening
metric to investigate functional connectivity at the source level, The proposed method was applied to dynamic functional
has been well validated in a previous study [29]. Therefore, networks constructed from the naturalistic music listening
we would not test the performance of connectivity metrics EEG dataset to detect the community structures across sub-
repeatedly in the current study. We only validated the ability jects. Fig. 3 shows the estimated brain network with spe-
of rank-(Lr , Lr , 1) BTD, for dynamic functional networks, to cific modular patterns from BTD components: their modular
extract the community structures and the temporal evolving structure profiles and their modulation scores by five musical
over time scales of minutes. features. The mean and standard deviation of the modulation
We constructed dynamic functional networks with adja- score were reported across subjects. Here, R = 5 components
cency matrices (n = 68 nodes, τ = 200 time points). The (with 4 community clusters for each, that is, Lr = 4, r =
tensor representation of the networks was obtained by the outer 1, ..., R) were extracted by BTD based on data fitting analysis
product of adjacency matrices
PR and temporal profiles. That is, (see Fig. 4), and we presented 3 components that showed
Msim = Ssim +Nsim = r=1 (Arsim Arsim T )◦crsim +Nsim , significant musical feature modulation. We observed unilateral
where Nsim ∈ Rn×n×τ is a noise tensor with dimensions auditory modular subnetworks (modules 2 and 4 in Row I of
same as Ssim . We predefined three community structures Fig. 3) and two bilateral frontoparietal functional subnetworks
(R = 3) and each of them included four communities or (modules 1 and 3 in Row I of Fig. 3). The auditory sub-
node clusters (Lr = 4, r = 1, 2, 3). We generated binary networks showed strong clustering in the temporal lobe. The
networks by using the node clusters. Their temporal evolution regions involved by the frontoparietal subnetworks were part
was modulated by triangle, square, and sine waves (Fig. 2). We of the frontoparietal network (FPN), which here was composed

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Transactions on Neural Systems and Rehabilitation Engineering. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2023.3277509
ZHU et al.: DYNAMIC COMMUNITY DETECTION WITH BLOCK COMPONENT ANALYSIS 7

Fig. 3. Results from music-listening data. A. For each subject, the modulation scores are estimated from the correlation analysis of temporal
courses of BTD components and music features (see Section II-G). Error bars display the standard errors of the mean across subjects. An asterisk
shows that the BTD component is modulated significantly differently (p < 0.05; corrected) from the surrogate data. B. The modular patterns of
3D visualization. Each dot/node indicates one brain region of the DKA atlas and nodes in the same community have the same color. The node
cluster or community is obtained from the Ar matrix with rank Lr , which encodes the node membership information. Row I shows two unilateral
auditory modules, sensorimotor and frontoparietal modules; row II indicates strong frontotemporal and temporoparietal modules; row III shows
frontotemporal and frontoparietal subnetworks.

of the dorsolateral prefrontal cortex and posterior parietal previous literature [27]. The involved regions were part of the
cortex. Such community structures were temporally modulated default mode network (DMN) which here contains temporal
by the Mode and Pulse Clarity features. Row II of Fig. 3 poles, the ventromedial prefrontal cortex, and the posterior
showed the sensorimotor networks (module 4), frontotemporal cingulate cortex. They also involved Broca’s area which was
subnetworks (modules 1 and 2), and temporoparietal subnet- often associated with semantic integration. The time course
works (module 3), which seemed to be related to the anterior of this modular pattern was significantly modulated by the
higher-order cognitive brain networks in accordance with Fluctuation Centroid. There were three modules detected in

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Transactions on Neural Systems and Rehabilitation Engineering. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2023.3277509
8 GENERIC COLORIZED JOURNAL, VOL. XX, NO. XX, XXXX 2023

Row III, showing strong clusters in the visual subnetworks, We also observed bilateral frontoparietal communities
frontotemporal and frontoparietal subnetworks. This modular (modules 1 and 3 of Row I in Fig. 3) in the FPN network that
pattern was significantly modulated by the Pulse Clarity. subserved an integrative function between periphery commu-
nities in the left and the right hemisphere during auditory or
semantic comprehension and unilateral auditory subnetworks
IV. D ISCUSSION
(modules 1 and 3 of Row I). This asymmetry modules 2 and
In this study, we proposed a BTD-based framework applied 4 might be associated with the language network that displays
to EEG data, which enabled us to characterize the dynamic some degree of hemispheric lateralization. The nodes of the
topological properties of electrophysiology brain networks language network would be implicated during naturalistic
during natural music listening. We formed a three-way tensor language comprehension task performance. Indeed, this left
including temporal evolution of functional connectivity at lateralized subnetwork is anchored in the angular gyrus with
the source level and then applied multilinear rank-(Lr , Lr , 1) extensions to the inferior frontal gyrus, inferior temporal
decomposition to detect the modular structures of time-varying gyrus, and a number of nodes spanning the inferior to superior
brain networks. We derived large-scale brain network topo- precentral gyrus. These regions are consistent with previous
logical structures during freely listening to music, which was accounts of semantic cognition [61], [62]. The parietal module
characterized by BTD components. Such BTD component, (module 4 of Row II in Fig. 3) was related to the motor
we referred to as a modular pattern, was represented with networks and it was believed that perception and execution of
a distinct topological pattern of functional networks across actions are strongly coupled in the brain as a result of learning
the set of predefined atlas regions spanning the whole brain. a sensorimotor task, which facilitated not only predicting
These modular patterns of topology-specific envelop-coupling the action of others but also interacting with them [63].
were found to be temporally modulated by musical features During music listening, a tight coupling emerged between
and corresponded to plausible brain functional sub-systems, the perception and production of sequential information in
consisting of auditory, sensorimotor, and higher-order cog- hierarchical organization [9], [63]. Brain regions associated
nitive subnetworks. To the authors’ knowledge, this might with motor networks could be involved due to imitation
be the first complete formulation of a BTD-based generative and synchronization during musical activities (e.g. ensemble
model method for module detection of electrophysiology brain playing or singing). These subnetworks involved in auditory
networks using ongoing EEG. areas (Row III in Fig. 3) played an important function in music
Simulation results showed the effectiveness of the proposed perception in agreement with previous studies [9], [50].
method for the detection of community structure in time- Tensor decomposition especially with CP model analysis
varying brain functional networks. When applied to the EEG methods has been applied for the multi-way neuroimaging
data recorded during the continuous music listening task, data in cognitive research since it enables multi-timescale
the proposed approach identified more diverse community dimensionality reduction both within and across subjects or
structures in addition to the typical assortative organization conditions with unsupervised learning [9], [39], [48]. This
in brain networks, which was related to auditory and semantic provides the possibility of module detection of dynamic time-
information processing as well as higher-order cognitive func- varying brain functional networks. The majority of studies for
tions. These types of network architecture seemed to be asso- brain networks typically applied the CP model to examine
ciated with music perception and were temporally modulated the temporal, spectral, or spatial features of brain connectivity
by the acoustic musical feature extracted from music. Their networks, which is unable to detect the modular structure in
topological structures might allow the network to engage in resulting CP components simultaneously. This results in the
a wider functional repertoire, e.g., integration of information requirement for further analysis such as clustering of the net-
across different brain regions in higher-order cognitive pro- work factor of CP components when looking at the topological
cesses. Different from CANDECOMP/PARAFAC(CP)-based structure of the networks. To overcome this limitation, the
methods, our proposed method allowed the mutual existence proposed methods adopted block component analysis, rank-
of multiple communities (modules or subnetworks) in the same (Lr , Lr , 1) BTD. The resulting BTD components include the
components (brain patterns) due to the multi-linear rank de- node factor matrix with rank-Lr , which is able to characterize
composition. For example, we found bilateral frontotemporal the topological structures of the latent network pattern. That
communities (modules 2 and 4 of Row II in Fig. 3) involved is the rank-Lr encodes Lr communities or node clusters.
a subdivision of the DMN network that subserves a semantic Actually, tensor decomposition with the CP model can be
integration and the ventromedial prefrontal cortex was typi- considered a special case of rank-(Lr , Lr , 1) block term de-
cally specialized for emotion regulation. The subnetworks also composition (i.e., Lr = 1). Intuitively, the rank Lr of the node
involved Broca’s areas, which are typically related to language factor matrix is capable of characterizing the membership of
processing. Previous studies demonstrated that brain functional nodes in the network. This, on the other hand, explains why
networks engaged in music processing have strict similarities the tensor decomposition with the CP model fails to detect the
with that of language processing [61], [62]. Thus, the nodes community structures (Lr = 1 one community left).
of the subnetwork including Broca’s areas could be implicated The key parameters for the BTD-based methods are the
during continuously listening to music. These subnetworks determination of the component number and rank number,
were also identified in the previous study with two independent which is less well prescribed and not a limitation of the
CP components based on CP-based methods [9]. proposed approach directly. In the absence of theoretically

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Transactions on Neural Systems and Rehabilitation Engineering. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2023.3277509
ZHU et al.: DYNAMIC COMMUNITY DETECTION WITH BLOCK COMPONENT ANALYSIS 9

R EFERENCES

[1] E. Bullmore and O. Sporns, “Complex brain networks: graph theoretical

analysis of structural and functional systems,” Nature reviews neuro-
science, vol. 10, no. 3, pp. 186–198, 2009.
[2] K. J. Friston, “Functional and effective connectivity in neuroimaging: a
synthesis,” Human brain mapping, vol. 2, no. 1-2, pp. 56–78, 1994.
[3] F. De Pasquale, S. Della Penna, A. Z. Snyder, C. Lewis, D. Mantini,
L. Marzetti, P. Belardinelli, L. Ciancetta, V. Pizzella, G. L. Romani,
et al., “Temporal dynamics of spontaneous meg activity in brain net-
works,” Proceedings of the National Academy of Sciences, vol. 107,
no. 13, pp. 6040–6045, 2010.
[4] M. J. Brookes, M. Woolrich, H. Luckhoo, D. Price, J. R. Hale, M. C.
Stephenson, G. R. Barnes, S. M. Smith, and P. G. Morris, “Investigating
Fig. 4. Data fitting curves in the function of the component number the electrophysiological basis of resting state networks using magne-
and rank number. A. Fit versus component number R obtained by toencephalography,” Proceedings of the National Academy of Sciences,
computing the BTD of EEG data for Lr = 4, r = 1, ..., R. B. Fit vol. 108, no. 40, pp. 16783–16788, 2011.
versus rank number Lr obtained by computing the BTD of EEG data for [5] B. He, L. Astolfi, P. A. Valdés-Sosa, D. Marinazzo, S. O. Palva, C.-G.
R = 5. Bénar, C. M. Michel, and T. Koenig, “Electrophysiological brain con-
nectivity: theory and implementation,” IEEE Transactions on Biomedical
Engineering, vol. 66, no. 7, pp. 2115–2137, 2019.
motivated methods for parameter selection, we here opted [6] R. M. Hutchison, T. Womelsdorf, E. A. Allen, P. A. Bandettini, V. D.
Calhoun, M. Corbetta, S. Della Penna, J. H. Duyn, G. H. Glover,
instead to repeat the data fitting analysis for different values J. Gonzalez-Castillo, et al., “Dynamic functional connectivity: promise,
to select a relatively rational number. Fig. 4A shows the data issues, and interpretations,” NeuroImage, vol. 80, pp. 360–378, 2013.
fitting with different component numbers when fixing the rank [7] D. Vidaurre, R. Abeysuriya, R. Becker, A. J. Quinn, F. Alfaro-Almagro,
Lr = 4. We can see the data fitting no longer increases or S. M. Smith, and M. W. Woolrich, “Discovering dynamic brain networks
from big data in rest and task,” NeuroImage, vol. 180, pp. 646–656,
increases slowly when R = 5. Fig. 4B shows the data fitting 2018.
with different rank numbers, indicating that data fitting almost [8] A. Kabbara, M. Khalil, G. O’Neill, K. Dujardin, Y. El Traboulsi,
stays constant as Lr increases. Finally, we set the R = 5 and F. Wendling, and M. Hassan, “Detecting modular brain states in rest
and task,” Network Neuroscience, vol. 3, no. 3, pp. 878–901, 2019.
Lr = 4 according to the data fitting analysis and previous [9] Y. Zhu, J. Liu, K. Mathiak, T. Ristaniemi, and F. Cong, “Deriving
experience. Note that such data fitting analysis only provides a electrophysiological brain network connectivity via tensor component
reference and instruction and is not able to accurately estimate analysis during freely listening to music,” IEEE Transactions on Neural
Systems and Rehabilitation Engineering, vol. 28, no. 2, pp. 409–418,
the underlying true numbers of BTD components. 2019.
For the parcellation, we chose the DK atlas as the template [10] Y. Zhu, J. Liu, T. Ristaniemi, and F. Cong, “Distinct patterns of
since the scalp electrodes are not very dense. Although beyond functional connectivity during the comprehension of natural, narrative
speech,” International journal of neural systems, vol. 30, no. 03,
the scope of the current work, the other atlas could also be p. 2050007, 2020.
used in our method after source leakage correction. In addition, [11] C. J. Stam, “Modern network science of neurological disorders,” Nature
note that the topological structure with specific four modules Reviews Neuroscience, vol. 15, no. 10, pp. 683–695, 2014.
is a whole pattern with one corresponding temporal evolution [12] U. Braun, A. Schaefer, R. F. Betzel, H. Tost, A. Meyer-Lindenberg, and
D. S. Bassett, “From maps to multi-dimensional network mechanisms
instead of four independent module patterns. of mental disorders,” Neuron, vol. 97, no. 1, pp. 14–31, 2018.
[13] Y. Zhu, X. Wang, K. Mathiak, P. Toiviainen, T. Ristaniemi, J. Xu,
Y. Chang, and F. Cong, “Altered eeg oscillatory brain networks during
V. C ONCLUSION music-listening in major depression,” International Journal of Neural
We introduced a framework based on a latent network Systems, vol. 31, no. 03, p. 2150001, 2021.
[14] E. A. Allen, E. Damaraju, S. M. Plis, E. B. Erhardt, T. Eichele, and V. D.
generative model and related it to BTD for detecting dy- Calhoun, “Tracking whole-brain connectivity dynamics in the resting
namic community evolution in time-varying brain networks state,” Cerebral cortex, vol. 24, no. 3, pp. 663–676, 2014.
during continuous music listening. It allows us to identify the [15] A. P. Baker, M. J. Brookes, I. A. Rezek, S. M. Smith, T. Behrens, P. J. P.
Smith, and M. Woolrich, “Fast transient networks in spontaneous human
topological structures of dynamic brain networks and their brain activity,” Elife, vol. 3, p. e01867, 2014.
time evolution during naturalistic stimuli. The majority of [16] D. Vidaurre, S. M. Smith, and M. W. Woolrich, “Brain network dynam-
approaches for brain networks failed to reveal the topological ics are hierarchically organized in time,” Proceedings of the National
structures of time-varying networks. Here, we apply block Academy of Sciences, vol. 114, no. 48, pp. 12827–12832, 2017.
[17] G. C. O’Neill, P. Tewarie, D. Vidaurre, L. Liuzzi, M. W. Woolrich, and
component analysis, rank-(Lr , Lr , 1) BTD, to the adjacent M. J. Brookes, “Dynamics of large-scale electrophysiological networks:
tensor. The node factor matrix of BTD components with rank- A technical review,” NeuroImage, vol. 180, pp. 559–576, 2018.
Lr is able to characterize the topological structures of the [18] C.-M. Ting, S. B. Samdin, M. Tang, and H. Ombao, “Detecting dynamic
community structure in functional brain networks across individuals: A
latent network pattern, which encodes Lr communities or node multilayer approach,” IEEE Transactions on Medical Imaging, vol. 40,
clusters. We validate the proposed method in simulation and no. 2, pp. 468–480, 2020.
then apply it to the EEG data recorded during free music [19] R. F. Betzel, L. Byrge, Y. He, J. Goñi, X.-N. Zuo, and O. Sporns,
“Changes in structural and functional connectivity among resting-state
listening. The identified brain patterns with distinct topological networks across the human lifespan,” NeuroImage, vol. 102, pp. 345–
structures were in line with those previously published in the 357, 2014.
fMRI and EEG studies. The proposed method looks valuable [20] O. Sporns and R. F. Betzel, “Modular brain networks,” Annual review
for the characterization of the temporal evolution of brain of psychology, vol. 67, pp. 613–640, 2016.
[21] L. E. Suárez, R. D. Markello, R. F. Betzel, and B. Misic, “Linking
networks with specific community structures during freely structure and function in macroscale brain networks,” Trends in Cogni-
listening to music or other naturalistic stimuli. tive Sciences, vol. 24, no. 4, pp. 302–315, 2020.

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Transactions on Neural Systems and Rehabilitation Engineering. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/TNSRE.2023.3277509
10 GENERIC COLORIZED JOURNAL, VOL. XX, NO. XX, XXXX 2023

[22] M. E. Newman and M. Girvan, “Finding and evaluating community [42] M. Roald, S. Bhinge, C. Jia, V. Calhoun, T. Adalı, and E. Acar,
structure in networks,” Physical review E, vol. 69, no. 2, p. 026113, “Tracing network evolution using the parafac2 model,” in ICASSP 2020-
2004. 2020 IEEE International Conference on Acoustics, Speech and Signal
[23] D. S. Bassett, N. F. Wymbs, M. A. Porter, P. J. Mucha, J. M. Carlson, and Processing (ICASSP), pp. 1100–1104, IEEE, 2020.
S. T. Grafton, “Dynamic reconfiguration of human brain networks during [43] L. Gauvin, A. Panisson, and C. Cattuto, “Detecting the community
learning,” Proceedings of the National Academy of Sciences, vol. 108, structure and activity patterns of temporal networks: a non-negative
no. 18, pp. 7641–7646, 2011. tensor factorization approach,” PloS one, vol. 9, no. 1, p. e86028, 2014.
[24] J. Ou, L. Xie, C. Jin, X. Li, D. Zhu, R. Jiang, Y. Chen, J. Zhang, L. Li, [44] L. De Lathauwer, “Decompositions of a higher-order tensor in block
and T. Liu, “Characterizing and differentiating brain state dynamics via terms—part ii: Definitions and uniqueness,” SIAM Journal on Matrix
hidden markov models,” Brain topography, vol. 28, no. 5, pp. 666–679, Analysis and Applications, vol. 30, no. 3, pp. 1033–1066, 2008.
2015. [45] G. Zhou, A. Cichocki, Q. Zhao, and S. Xie, “Nonnegative matrix
[25] S. Ma, V. D. Calhoun, R. Phlypo, and T. Adalı, “Dynamic changes and tensor factorizations: An algorithmic perspective,” IEEE Signal
of spatial functional network connectivity in healthy individuals and Processing Magazine, vol. 31, no. 3, pp. 54–65, 2014.
schizophrenia patients using independent vector analysis,” NeuroImage, [46] A. Cichocki, D. Mandic, L. De Lathauwer, G. Zhou, Q. Zhao, C. Caiafa,
vol. 90, pp. 196–206, 2014. and H. A. Phan, “Tensor decompositions for signal processing applica-
[26] D. Vidaurre, A. J. Quinn, A. P. Baker, D. Dupret, A. Tejero-Cantero, tions: From two-way to multiway component analysis,” IEEE signal
and M. W. Woolrich, “Spectrally resolved fast transient brain states in processing magazine, vol. 32, no. 2, pp. 145–163, 2015.
electrophysiological data,” NeuroImage, vol. 126, pp. 81–95, 2016. [47] N. D. Sidiropoulos, L. De Lathauwer, X. Fu, K. Huang, E. E. Papalex-
[27] D. Vidaurre, L. T. Hunt, A. J. Quinn, B. A. Hunt, M. J. Brookes, akis, and C. Faloutsos, “Tensor decomposition for signal processing and
A. C. Nobre, and M. W. Woolrich, “Spontaneous cortical activity machine learning,” IEEE Transactions on Signal Processing, vol. 65,
transiently organises into frequency specific phase-coupling networks,” no. 13, pp. 3551–3582, 2017.
Nature communications, vol. 9, no. 1, pp. 1–13, 2018. [48] F. Cong, V. Alluri, A. K. Nandi, P. Toiviainen, R. Fa, B. Abu-Jamous,
[28] N. Leonardi, J. Richiardi, M. Gschwind, S. Simioni, J.-M. Annoni, L. Gong, B. G. Craenen, H. Poikonen, M. Huotilainen, et al., “Linking
M. Schluep, P. Vuilleumier, and D. Van De Ville, “Principal components brain responses to naturalistic music through analysis of ongoing eeg
of functional connectivity: a new approach to study dynamic brain and stimulus features,” IEEE Transactions on Multimedia, vol. 15, no. 5,
connectivity during rest,” NeuroImage, vol. 83, pp. 937–950, 2013. pp. 1060–1069, 2013.
[29] G. C. O’Neill, P. K. Tewarie, G. L. Colclough, L. E. Gascoyne, B. A. [49] L. De Lathauwer and D. Nion, “Decompositions of a higher-order tensor
Hunt, P. G. Morris, M. W. Woolrich, and M. J. Brookes, “Measurement in block terms—part iii: Alternating least squares algorithms,” SIAM
of dynamic task related functional networks using meg,” NeuroImage, journal on Matrix Analysis and Applications, vol. 30, no. 3, pp. 1067–
vol. 146, pp. 667–678, 2017. 1083, 2008.
[30] P. J. Mucha, T. Richardson, K. Macon, M. A. Porter, and J.-P. Onnela, [50] V. Alluri, P. Toiviainen, I. P. Jääskeläinen, E. Glerean, M. Sams,
“Community structure in time-dependent, multiscale, and multiplex and E. Brattico, “Large-scale brain networks emerge from dynamic
networks,” science, vol. 328, no. 5980, pp. 876–878, 2010. processing of musical timbre, key and rhythm,” NeuroImage, vol. 59,
no. 4, pp. 3677–3689, 2012.
[31] E. Al-Sharoa, M. Al-Khassaweneh, and S. Aviyente, “Tensor based
[51] Y. Zhu, C. Zhang, H. Poikonen, P. Toiviainen, M. Huotilainen, K. Math-
temporal and multilayer community detection for studying brain dy-
iak, T. Ristaniemi, and F. Cong, “Exploring frequency-dependent brain
namics during resting state fmri,” IEEE Transactions on Biomedical
networks from ongoing eeg using spatial ica during music listening,”
Engineering, vol. 66, no. 3, pp. 695–709, 2018.
Brain topography, vol. 33, no. 3, pp. 289–302, 2020.
[32] L.-E. Martinet, M. Kramer, W. Viles, L. Perkins, E. Spencer, C. Chu,
[52] A. Hyvärinen and E. Oja, “Independent component analysis: algorithms
S. Cash, and E. Kolaczyk, “Robust dynamic community detection with
and applications,” Neural networks, vol. 13, no. 4-5, pp. 411–430, 2000.
applications to human brain functional networks,” Nature communica-
[53] F. Tadel, S. Baillet, J. C. Mosher, D. Pantazis, and R. M. Leahy, “Brain-
tions, vol. 11, no. 1, pp. 1–13, 2020.
storm: a user-friendly application for meg/eeg analysis,” Computational
[33] J. M. Mueller, L. Pritschet, T. Santander, C. M. Taylor, S. T. Grafton, intelligence and neuroscience, vol. 2011, 2011.
E. G. Jacobs, and J. M. Carlson, “Dynamic community detection reveals [54] F.-H. Lin, J. W. Belliveau, A. M. Dale, and M. S. Hämäläinen, “Dis-
transient reorganization of functional brain networks across a female tributed current estimates using cortical orientation constraints,” Human
menstrual cycle,” Network Neuroscience, vol. 5, no. 1, pp. 125–144, brain mapping, vol. 27, no. 1, pp. 1–13, 2006.
2021. [55] R. S. Desikan, F. Ségonne, B. Fischl, B. T. Quinn, B. C. Dickerson,
[34] K. Tu, B. Ribeiro, A. Swami, and D. Towsley, “Temporal cluster- D. Blacker, R. L. Buckner, A. M. Dale, R. P. Maguire, B. T. Hyman,
ing in dynamic networks with tensor decomposition,” arXiv preprint et al., “An automated labeling system for subdividing the human cerebral
arXiv:1605.08074, 2016. cortex on mri scans into gyral based regions of interest,” NeuroImage,
[35] Z. Chen, C. Chen, Z. Zheng, and Y. Zhu, “Tensor decomposition for vol. 31, no. 3, pp. 968–980, 2006.
multilayer networks clustering,” in Proceedings of the AAAI Conference [56] J. F. Hipp, D. J. Hawellek, M. Corbetta, M. Siegel, and A. K. Engel,
on Artificial Intelligence, vol. 33, pp. 3371–3378, 2019. “Large-scale cortical correlation structure of spontaneous oscillatory
[36] E. Al-sharoa, M. A. Al-khassaweneh, and S. Aviyente, “Detecting and activity,” Nature neuroscience, vol. 15, no. 6, pp. 884–890, 2012.
tracking community structure in temporal networks: A low-rank + sparse [57] A. Cichocki, R. Zdunek, A. H. Phan, and S.-i. Amari, Nonnegative
estimation based evolutionary clustering approach,” IEEE Transactions matrix and tensor factorizations: applications to exploratory multi-way
on Signal and Information Processing over Networks, vol. 5, no. 4, data analysis and blind source separation. John Wiley & Sons, 2009.
pp. 723–738, 2019. [58] N. Vervliet, O. Debals, L. Sorber, M. Van Barel, and L. De Lathauwer,
[37] G. Frusque, J. Jung, P. Borgnat, and P. Gonçalves, “Multiplex network “Tensorlab 3.0,” Mar. 2016. Available online.
inference with sparse tensor decomposition for functional connectivity,” [59] Y. Cheng, M. Riesmeyer, J. Haueisen, and M. Haardt, “Using the multi-
IEEE transactions on Signal and Information Processing over Networks, linear rank-(lr, lr, 1) decomposition for the detection of the 200 hz band
vol. 6, pp. 316–328, 2020. activity in somatosensory evoked magnetic fields and somatosensory
[38] Y. Zhu, X. Li, T. Ristaniemi, and F. Cong, “Measuring the task induced evoked electrical potentials,” IEEE Access, vol. 9, pp. 106232–106244,
oscillatory brain activity using tensor decomposition,” in ICASSP 2019- 2021.
2019 IEEE International Conference on Acoustics, Speech and Signal [60] D. Prichard and J. Theiler, “Generating surrogate data for time series
Processing (ICASSP), pp. 8593–8597, IEEE, 2019. with several simultaneously measured variables,” Physical review letters,
[39] Y. Zhu, J. Liu, C. Ye, K. Mathiak, P. Astikainen, T. Ristaniemi, and vol. 73, no. 7, p. 951, 1994.
F. Cong, “Discovering dynamic task-modulated functional networks with [61] L. Fadiga, L. Craighero, and A. D’Ausilio, “Broca’s area in language,
specific spectral modes using meg,” NeuroImage, vol. 218, p. 116924, action, and music,” Annals of the New York academy of sciences,
2020. vol. 1169, no. 1, pp. 448–458, 2009.
[40] J. Liu, Y. Zhu, H. Sun, T. Ristaniemi, and F. Cong, “Sustaining attention [62] S. Koelsch, E. Kasper, D. Sammler, K. Schulze, T. Gunter, and A. D.
for a prolonged duration affects dynamic organizations of frequency- Friederici, “Music, language and meaning: brain signatures of semantic
specific functional connectivity,” Brain Topography, vol. 33, pp. 677– processing,” Nature neuroscience, vol. 7, no. 3, pp. 302–307, 2004.
692, 2020. [63] V. Alluri, P. Toiviainen, I. Burunat, M. Kliuchko, P. Vuust, and E. Brat-
[41] E. Acar, M. Roald, K. M. Hossain, V. D. Calhoun, and T. Adali, “Tracing tico, “Connectivity patterns during music listening: Evidence for action-
evolving networks using tensor factorizations vs. ica-based approaches,” based processing in musicians,” Human Brain Mapping, vol. 38, no. 6,
Frontiers in neuroscience, vol. 16, 2022. pp. 2955–2970, 2017.

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/