Multilayer Networks: For Modeling and Analysis of Big Data
Abhishek Santra, Hafsa Billah and Sharma Chakravarthy
Information Technology Lab, CSE Department, University of Texas at Arlington, Texas, U.S.A.
Keywords:
Multilayer Networks, Modeling, Analysis, Big Data.
Abstract:
In this position paper, we make a case for the appropriateness, utility, and effectiveness of graph models for
big data analysis focusing on Multilayer Networks (or MLNs) – a specific type of graph. MLNs have been
shown to be more appropriate for modeling complex data compared to their traditional counterparts. MLNs
have also been shown to be useful for diverse data types, such as videos and information integration. Further,
MLNs have been shown to be flexible for computing analysis objectives from diverse application domains
using extant and new algorithms. There is research for automating the modeling of MLNs using widely used
EER (Enhanced/Extended Entity Relationship) or Unified Modeling Language (UML) approaches.
We start by discussing different graph models and their benefits and limitations. We demonstrate how MLNs
can be effectively used to model applications with complex data. We also summarize the work on the use of
EER models to generate MLNs in a principled manner. We elaborate on analysis alternatives provided by
MLNs and their ability to match analysis needs. We show the use of MLNs for - i) traditional data analysis, ii)
video content analysis, iii) complex data analysis, and iv) propose the use of MLNs for information integration
or fusion. We show examples drawn from the literature of their modeling and analysis usage. We conclude that
graphs, specifically MLNs provide a rich alternative to model and analyze big data. Of course, this certainly
does not preclude newer data models that are likely to come along.
1 INTRODUCTION
Big data analytics is predicated upon our ability to
model and analyze disparate, complex data sets and
associated application objectives. Relational and
object-oriented data models have served well for
modeling and analyzing transactional data sets that
need to be managed over long periods. NoSQL data
models filled the gap in modeling and analysis for
data sets for which earlier data models were not best
suited. New data models including graph models are
gaining importance due to the diverse types of social
networks and other data types being used for mining,
knowledge discovery, querying, and analysis.
Figure 1: Life Cycle Flow Chart of Mining.
In this paper, we focus on the applicability and
versatility of graphs, especially Multilayer Networks
(MLNs) for moving towards modeling and analysis of
big data. In contrast to the mining approach shown in
Figure 1, big data analysis needs to be addressed us-
ing a life cycle starting from modeling to drill-down
and visualization. Currently, graph models are gen-
erated manually for a given data set without using
any principled approach. For many data sets, both
modeling and analysis computations are quite differ-
ent from the ones addressed in earlier data models. In
this paper, instead of generating a schema, application
requirements and data are transformed into different
types of graphs including MLNs. Moreover, an anal-
ysis may require graph computations, such as short-
est path, substructure discovery, community, central-
ity (e.g., hubs), or their combination. Once the cho-
sen data model is generated and the objectives are
mapped into appropriate computations, any available
package/algorithm can be used. Finally, the analysis
results need to be drilled down and visualized in mul-
tiple ways for decision-making and for taking action.
We present several results from the literature to con-
vince the reader that this workflow is needed. Figure
2 shows our view of the big data analysis life cycle
from gathered application requirements to analysis
of objectives to result drill-down with visualization.
Only graph and MLN models are shown. This work-
flow is iterative.
Drill-down of analysis results is critical, espe-
cially for diverse data that has both structure and se-
Santra, A., Billah, H. and Chakravarthy, S.
Multilayer Networks: For Modeling and Analysis of Big Data.
DOI: 10.5220/0012997200003838
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 16th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2024) - Volume 1: KDIR, pages 383-390
ISBN: 978-989-758-716-0; ISSN: 2184-3228
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.
383

Figure 2: Life Cycle of Big Data Modeling and Analysis
using Graphs and MLNs.
mantics. For example, it is not sufficient to know
the objects in a community, but additional object de-
tails are needed, similarly, for a centrality hub. For
graph and MLN models, we also need to know the
edges within and across layers, if any. From a com-
putation/efficiency perspective, minimal information
needs to be used for analysis whereas the drill-down
phase needs to expand upon to the desired extent.
Visualization is not new either and there exists a
wide variety of tools for visualizing base data, results,
and drilled-down information in multiple ways. Sev-
eral data visualization platforms are available (GeP,
2014, Samant et al., 2021). Due to space constraints,
we will not discuss drill-down and visualization in
this paper. The contributions of this paper are:
• Complete Life cycle for big data analytics in
comparison with mining
• Graph and MLN models, and analysis alterna-
tives
• Use and applicability of MLNs for complex data
• Graphs and MLNs applicability for video data
analysis
• MLNs Applicability for information integra-
tion/fusion
The rest of the paper is organized as follows. Rel-
evant literature and different graph models are dis-
cussed in Sec. 2 and Sec. 3 respectively. The use of
MLNs to model and analyze complex data is summa-
rized in Sec. 4. The use of MLN in lieu of graphs
is discussed in Sec. 5. Graphs and MLNs applicabil-
ity to model and analyze video data is summarized in
Sec. 6. Finally, the use and applicability of MLNs for
information integration/fusion are discussed in Sec. 7.
We conclude and outline future work in Sec. 8.
2 RELATED WORK
We discuss here how different phases of the lifecycle
have been addressed in the literature.
EER Modeling: Since the 70s, EER model (Chen,
1976) has served as a methodology for database de-
sign, by representing data and functionality require-
ments of real-world applications in a precise manner
by identifying entities, attributes, and relationships
among them. However, with the emergence of data
sets with multiple entity types and relationships along
with complex analysis requirements, such as shortest
paths, important neighborhoods, dominant nodes (or
groups of nodes), etc., the relational data model was
not adequate for modeling and analysis. Recently,
there has been some work in modeling graphs from
EER diagrams but is limited to simple and attributed
graphs only (Roy-Hubara et al., 2017, Angles, 2018).
Graph and MLN Models: When a graph is used as
a data model, the choice of nodes, edges, and their la-
bels becomes important. There are multiple ways of
creating them depending on the analysis objectives.
Further, creating edges needs similarity/proximity cri-
teria which need to be specified/identified. There
needs to be a systematic and configurable approach
for converting raw data sets (.csv files, extracted video
contents, etc.) to graphs or MLN layers. Only
recently, there has been some work (Komar et al.,
2020, Santra et al., 2022) on extending the EER ap-
proach to generate MLN models.
Graph and MLN Analysis: There is substan-
tial work in the area of simple, attributed graphs
and MLNs. For simple graphs, many algorithms
have been developed for shortest paths, spanning
trees, community detection, centrality measures, and
cliques. The breadth and depth-first approaches are
also used for many algorithms. For attributed graphs,
substructure discovery (Holder et al., 1994, Padman-
abhan and Chakravarthy, 2009, Yan and Han, 2002)
for interesting exact and inexact or similar substruc-
tures, and graph search and querying (Das et al.,
2020) have been developed. For MLNs, algorithms
have been developed for homogeneous (HoMLN) and
heterogeneous (HeMLN) MLNs. Community detec-
tion algorithms have been extended to HoMLNs (re-
view: (Kim and Lee, 2015, Magnani et al., 2021)).
Further, methods have been developed to deter-
mine centrality measures to identify highly influ-
ential nodes (Sol
´
e-Ribalta et al., 2014, Zhan et al.,
2015). Recently developed decoupling-based ap-
proaches combine partial analysis results from indi-
vidual layers systematically in a loss-less manner to
compute communities (Santra et al., 2017) or cen-
trality hubs (Pavel et al., 2023) for layer combina-
tions. Majority of HeMLN work (reviews in (Shi
et al., 2017, Sun and Han, 2013)) focuses on de-
veloping meta-path based methods for object simi-
larity, object classification, missing link prediction,
ranking/co-ranking, and recommendations. Few ex-
isting works generate clusters of entities (Melamed,
2014). Most of them concentrate mainly on inter-
layer edges and not the networks themselves.
KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval
384

Graph Models for Video Analysis: Several custom
approaches have been developed for modeling videos
as scene graphs (Ji et al., 2020, Ou et al., 2022) by
training deep learning algorithms and can perform
fixed types of analysis (Billah et al., 2024). They
need to be retrained or a new algorithm is required to
perform a new type of analysis. Several frameworks
are also available which models extracted video con-
tents as attributed graphs (Yadav et al., 2020, Zhang
et al., 2023) and perform analysis on them. How-
ever, they do not consider all the extracted video con-
tents for modeling and only support simple analysis
such as counting the number of objects. They can-
not perform complex analyses (e.g., finding groups)
on videos. Graphs and MLNs can be leveraged to
model all the extracted video contents and new al-
gorithms/operators need to be developed to perform
interesting analysis on videos.
3 GRAPHS FOR BIG DATA
ANALYSIS
Graphs capture relationships between entities in ap-
plication data using nodes and edges. This represen-
tation allows us to perform various analyses based on
the graph structure and relationships found in the data.
3.1 Graph Types Used as Data Models
A simple graph is defined as (V, E) where V is a set
of vertices or nodes and E is a set of edges connect-
ing two distinct vertices. E is a subset of V × V. The
edges are assumed to be unweighted, either directed
or undirected, and loops and multiple edges between
nodes are not allowed. Typically, vertices have unique
numbers, but labels of nodes and edges need not be
unique. These graph models are widely used for mod-
eling and analyzing applications.
An attributed graph (also called a multigraph)
is defined as (V, E, φ) where V is a set of vertices
or nodes, E is a set of edges connecting two dis-
tinct vertices, and φ is a function mapping of E to
{{x, y} | x, y ∈ V and x ̸= y}. If the distinctness of
nodes is removed, loops will be allowed as well. The
main advantage of a multigraph or attributed graph
from a modeling viewpoint is that it captures multi-
ple entities and multiple relationships between enti-
ties. Multiple labels can be associated with nodes and
entities. With the attributed graph model, it is possible
to include relevant information from the data descrip-
tion as labels and hence is more expressive as a model
than a simple graph model.
An MLN is a network of simple graphs (or
forests). In this model, every layer represents a dis-
tinct relationship among entities with respect to a sin-
gle (or combination of) feature(s). The sets of entities
across layers, which may or may not be of the same
type, can be related to each other too.
An MLN can be used to separate entities and cor-
responding relationships from an attributed graph into
separate layers where each layer is a simple graph.
This provides more clarity in understanding and pro-
cessing. MLNs are widely used for modeling com-
plex data sets with multiple types of entities and mul-
tiple relationships between the same types of entities.
They can also capture relationships between different
types of entities.
Figure 3: Multilayer Network Types.
Based on the type of relationships and entities,
MLNs can be classified into three types. Layers of
a homogeneous MLN (HoMLN) are used to model
different relationships among the same entity types
like movie actors who are linked based on co-acting
(i.e., they act together in a movie) or have simi-
lar average rating or have worked in similar genres
(Figure 3(a)). Thus, V
1
= V
2
= . . . = V
n
and inter-
layer edge sets are empty as no relations across lay-
ers are necessary. Relationships among different
types of entities like researchers (connected by co-
authorship), research papers (connected if published
in the same conference), and year (related by pre-
defined ranges/eras) are modeled through heteroge-
neous MLN (HeMLN) (Figure 3(b)). The inter-
layer edges represent the relationship across layers
like writes, published-in, and active-in. In addition to
being collaborators, researchers may be social media
friends. Thus, to model multi-feature data that cap-
ture multiple relationships within and across dif-
ferent types of entity sets, a combination of homoge-
neous and heterogeneous MLNs is used, termed hy-
brid MLN (HyMLN), as shown in Figure 3(c). Here,
the first and the third layer have the same node types
(researchers) linked to the city nodes they reside in,
which are in turn connected based on the flight net-
work (second layer).
Multilayer Networks: For Modeling and Analysis of Big Data
385

The above graph types and MLN variants pro-
vide alternatives for matching modeling and analysis
needed for application data. Further, MLNs provide
clarity in understanding the data set. Additionally, the
availability of algorithms for a specific graph model
also plays a key role in the choice of the graph model.
For instance, there are not many algorithms avail-
able for attributed graphs in contrast to simple graphs.
There is considerable ongoing research in developing
algorithms for the MLNs (Boden et al., 2012, Santra
et al., 2017) due to the clarity of the model. Hence,
MLNs are preferred for modeling complex data sets.
3.2 EER Modeling Extensions
In contrast to the relational data model, a principled
approach to convert application requirements into a
chosen graph model (simple, attributed, or MLN) is
lacking. However, recently there has been some work
in this regard (Komar et al., 2020, Santra et al., 2022)
leading to the wider use of MLNs. Broadly, the enti-
ties in the EER diagram dictate the formation of lay-
ers with the entity instances as layer nodes and the
binary self relationship defining the intra-layer edges.
The binary non-self relationships define the inter-
layer edges. Some relationships are self-explanatory
and can be easily mapped into edges like friendships,
siblings, direct flights, and so on. However, some re-
lationships are non-explicit like “two actors working
in similar genre of movies” for which the EER model
needs to have a parameter attribute for the relation-
ship that defines the similarity metric and threshold.
The value of these parameter attributes will be used
to generate the edges in the MLN. Currently, we are
developing algorithms for converting EER to any type
of graph, not just MLN. More research is needed in
this area to make analysis easier.
4 MLNs FOR BIG DATA
ANALYSIS
Depending on the analysis requirements, the Google
Knowledge Base (GKB) data set can be modeled as
different types of MLNs. For instance, there exist
multiple relationships among the same set of people
- whether they are married to each other or have the
same birth state or studied in the same university, and
so on. This gives rise to a homogeneous GKB MLN
with the same set of nodes being connected differently
in each layer (Figure 4(a)). Similarly, Figure 4(b)
shows an HeMLN where both layers have different
sets of entities - person, and company.
Figure 4: Google Knowledge Base modeled as MLNs.
The person nodes are connected if they studied
in the same university, the company nodes are con-
nected if they focus on similar fields, and the person
nodes are connected to the company nodes that they
founded/established through inter-layer edges. This
may also be extended to Hybrid MLNs if two differ-
ent person layers are connected to a company layer.
4.1 MLN: Multiple Analysis Choices
Figure 5 shows three MLN analysis alternatives. Fig-
ure 5(a) shows an MLN conflated into a simple graph
by aggregating layers. These aggregation approaches,
termed type-independent (Domenico et al., 2014) and
projection-based (Berenstein et al., 2016), ignore type
information. Hence, they do not support structure and
semantics preservation without elaborate mappings as
they aggregate or collapse layers into a simple graph
in different ways. As observed in the literature, with-
out additional mappings, currently-used aggregation
approaches are likely to result in some information
loss, distortion of properties, or hide the effect of dif-
ferent entity types and/or different intra- or inter-layer
relationships (Kivel
¨
a et al., 2013, De Domenico et al.,
2014). At the other end of the spectrum, Figure 5(c)
shows the same MLN layers and result computation
by traversing the MLN as is.
Figure 5: (a) Lossy Vs. (b) Decoupling Vs. (c) Whole
MLN approaches (Santra et al., 2022).
Figure 5(b) on the other hand proposes an ap-
proach, termed networking decoupling, where net-
work property for each layer is computed indepen-
KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval
386

dently (possibly in parallel) in the analysis (Ψ) phase
and compose them using a binary operator Θ. This
approach has been shown to be effective and can
be done using Boolean operations for HoMLNs and
HeMLNs without losing type information. Further-
more, it is more efficient than the approaches shown
in Figure 5(a) or (c). Finally, the clarity of modeling
using MLNs is retained as well.
5 USE OF MLNs IN LIEU OF
GRAPHS
Based on daily life interactions (education, social me-
dia platforms, restaurant check-ins, healthcare check-
up appointments, etc.) different facts are available on
the web. In terms of knowledge base, “different facts”
about a person are captured in the GKB. Freebase
captures such information for famous personalities:
birth place and residence, education institutions at-
tended, birth and death date (if available), companies
worked in/founded, family-based relationships and so
on (Bollacker et al., 2008). Here, people, universi-
ties, companies, and states are related to each other
based on explicitly available interactions or relation-
ships. Some interesting analysis objectives can be:
(GKB-O1) Find frequently occurring patterns
among states, based on university locations
and place of company headquarters for the
entrepreneurs.
(GKB-O2) Find groups of people who were born in
the same state and have studied in the same uni-
versity.
(GKB-O3) For each group of founders who have
studied in the same university, find out the most
popular focus field among the group of similar
companies that they have founded.
Although objective (GKB-O1) can be computed us-
ing traditional graph models, MLNs are needed for
objective (GKB-O2) and others similar to that. The
HoMLN shown in Figure 4(a) is required to address
(GKB-O2). In this case, we need to “Find groups
of people who were born in the same state and have
studied in the same university”. Here “grouping” key-
word means that we need to compute communities
among the people nodes, followed by AND compo-
sition (due to the “and” keyword). For AND com-
position, here the CE-AND composition algorithm is
used that intersects the community edges, then per-
form a connected component analysis to obtain the
group of nodes that are tightly connected in both the
layers (Santra et al., 2022, Santra et al., 2017). Thus,
the analysis expression based on the decoupling ap-
proach can be expressed as:
Expression: Ψ(PERSON-Born-in-same-state) Θ
Ψ(PERSON-Studied-in-same-university);
where Ψ = Community; Θ = CE-AND (composition)
Similarly, for (GKB-O3), the HeMLN shown in
Figure 4(b) is used. Here, “For each group of
founders who have studied in the same university,
we need to find out the most popular focus field
among the group of similar companies that they have
founded.” Thus, communities need to be detected
in both person and company layers, which become
meta-nodes in the bipartite graph. The number of
inter-layer edges between the constituent nodes of
each pair of meta nodes will define the edge weight.
Finally, maximal weighted matching (MWM) will
give us the required optimal pairing of person and
company communities (Santra et al., 2022). The anal-
ysis expression is as follows:
Expression: Ψ(PERSON-Studied-in-same-univer- sity)
Θ Ψ(COMPANY-Focus-on-similar-fields);
where Ψ = Community; Θ = MWM (bipartite maximum
weighted matching)
6 GRAPHS/MLNs FOR VIDEO
ANALYSIS
Our goal, as part of big data analysis, is to han-
dle different data types (4 Vs of big data) in the
same way we handle structured and tabular data. If
videos (or extracted contents) can be modeled using
graphs/MLNs, the same life cycle approach can be
applied for video analysis, enabling the modeling and
analysis of video data alongside other data types. As
discussed in Sec. 2, the existing custom approaches
for video analysis require new software/algorithm/re-
training to perform a new analysis. Hence, this ap-
proach does not lend itself to the holistic analysis re-
quired for big data. In contrast, if big data analy-
sis were to include video analysis in mainstream data
processing, a different approach would be needed.
Some works in the literature used graphs for video
analysis as explained in Sec. 2. Recently, (Billah
et al., 2024) proposed a novel approach for video
analysis that has the potential to advance big data
analysis to include videos. This approach is novel
as video contents are extracted once (using existing
Video Content Extraction (VCE) algorithms), mod-
eled, and then analyzed to identify a variety of situ-
ations from them. This approach has several advan-
tages: i) video contents are extracted only once, ii)
it is possible to model these extracted contents com-
pletely, iii) several analysis expressions can be formu-
Multilayer Networks: For Modeling and Analysis of Big Data
387

lated and computed on them, iv) both “ad hoc” and
“what if” analysis can be supported, and v) most im-
portantly, this can be extended for real-time analysis.
A workflow of open-source VCE algorithms can
be used for extracting object bounding boxes and
class labels (with a confidence score) using object de-
tection (YOLO (Wang et al., 2024)) algorithm, unique
identifier (object id) for each object and feature vec-
tors using object tracking (Bot-sort (Aharon et al.,
2022)) algorithm, and pose coordinates using pose es-
timation (HRNet (Wang et al., 2020)) algorithm.
Modeling of Extracted Video Contents: The dif-
ferent types of extracted video contents can be mod-
eled in multiple ways. Two promising models that are
being explored in the literature are the extended re-
lational model (Billah and Chakravarthy, 2024) and
the graph model (Billah et al., 2024). If it is mod-
eled using an extended relational model, Continuous
Query Language (CQL) (an extension of the widely-
used Structured Query Language (SQL)) can be used.
If the extracted contents are modeled as graphs, dif-
ferent graph analysis techniques can be used. The ra-
tionale for using multiple models is that some analysis
may be easier in one model as compared to the other.
For example, clustering of objects is easier using the
graph model than the extended relational model. We
will focus on the graph model as this paper is about
the utility of graphs and MLNs for big data analysis.
To represent extracted video contents as graphs,
nodes and edges need to be identified and other re-
lated information (e.g., the feature vectors, bounding
boxes, etc.) needs to be associated properly for com-
putation. Many analyses involve objects. Hence, ob-
jects are represented as nodes and Object id as node
id in the literature. There are multiple choices to cre-
ate edges (e.g., the distance between objects, and their
spatial relationship in a frame, etc.). Figure 6 shows
a graph representation of a sample video frame with
nodes with two labels: frame id ( f
id
) and object class
label (O
l
) and edges (based on the objects bounding
box centroid distance).
A spectrum of alternatives exists for the graph rep-
resentation, each with different advantages and disad-
vantages. It is possible to model the entire video as
Figure 6: Graph representation of a sample video frame.
one graph (model M
1
) using object id for nodes (with
a large amount of information with each node). It is
also possible to create a graph for each frame (model
M
F
) (shown in Figure 6), where the number of graphs
will be equal to the number of non-empty frames F in
the video. Options in-between are also possible where
a forest of g ( 1 ≤ g ≤ F) graphs (model M
g
) can
be generated by aggregating the consecutive frames
into a graph based on some constraints, with vary-
ing numbers of graphs for different videos. The in-
between alternatives allow us to compress node labels
and edges in different ways reducing the storage re-
quired and can also reduce computational complexity
as the graphs are generated in some logical manner.
Video Content Analysis Using the Graph Models:
Below, we indicate video analysis examples using
graph models from the literature.
1. Identifying Groups (Billah et al., 2024): In as-
sisted living environment videos, it is useful to
identify isolated individuals (not participating in
group discussions, etc.) This analysis has been re-
ported in (Billah et al., 2024) to cluster individuals
in video frames by leveraging K-Means clustering
on model M
F
where nodes are objects and edges
are the object bounding box centroid distances.
2. Identifying if a Parking Slot is Occupied (Ya-
dav et al., 2020): In surveillance videos, it is of-
ten important to know which parking spaces are
occupied. This analysis has been reported in (Ya-
dav et al., 2020) using model M
F
, where nodes are
objects and edges are spatial bounding box rela-
tionships (e.g., overlap, inside, etc.) between ob-
jects in a frame. Their proposed algorithm identi-
fies a parking lot as occupied if the parking lot and
a car’s bounding box overlap over a threshold.
In summary, extracted video contents are shown to be
modeled and analyzed using alternative graph mod-
els and analysis algorithms. MLNs come in handy
to model multiple graphs (or videos) as different lay-
ers and perform combined analysis. HoMLNs can be
used by connecting object ids from different graphs
generated from the same video or by connecting ob-
ject ids from different videos if their feature vectors
match. Once modeled appropriately, interesting anal-
ysis (e.g., groups of objects entering and exiting a
premise after n minutes of each other) can be per-
formed using graphs/MLNs.
7 MLNs FOR INFORMATION
FUSION
Analysis of a single modality/data type has been the
KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval
388

major focus until now, be it structured (e.g., stream
data processing (Barbieri et al., 2010)) or unstructured
data (e.g., image and video analysis (Zhang et al.,
2023, Yadav et al., 2020), text and natural language
processing (Otter et al., 2020)). Yet, when all or a sub-
set of these data types must be analyzed holistically,
several challenges emerge. These problems have been
categorized under various headings, such as data fu-
sion, multi-modal data analysis, and others which are
limited in scope and context (Atrey et al., 2010). The
challenges originate due to the lack of approaches that
can effectively perform information fusion both at the
modeling and analysis stages. Therefore, the holistic
approach needs to accommodate modeling, and anal-
ysis techniques for objectives for performing knowl-
edge discovery. In our view, MLNs with their model-
ing and analysis advantages provide a path to explore
information fusion. Many applications, such as cyber-
security, healthcare, and surveillance can benefit from
this. We illustrate this with an example.
Sample Application – Healthcare: Patient data is
collected in diverse formats by different specialists
over time. This data constitutes the patient’s medi-
cal records including demographics, hospital/doctor
visits, vital signs, medications, progress notes, aller-
gies, radiology images, and laboratory results, and
can be further enriched by exercise data, etc. This
data is both spatial and temporal. When all this data
is accumulated, holistic knowledge discovery over an
individual and the population is possible. This ap-
plication with big data characteristics can be used
for personalized care using querying, searching, and
mining. MLNs can be used for effectively modeling
this data and for flexible analysis. Layers that can
be identified are: i) Demographics Layer(s): Pa-
tient nodes are connected by edges based on demo-
graphics (age, ethnicity, profession, education level,
etc.), ii) Image/Video Layer(s): Patient nodes are
connected based on the similarity of patterns present
in them (X-rays, MRIs, EKG, and CT Scans), iii)
Pathology Layer(s): Patient nodes are connected
based on the similarity of indicators (e.g., high sugar,
high/low BP, etc.), iv) Vaccination Layer(s): Person
nodes are connected based on the number of doses
and type of shots. These layers can be generated for
county/city/state as needed.
Figure 7(a) illustrates 4 possible layers of the
hybrid healthcare MLN, with the inter-layer edges.
For example, the demographics layer can be linked
with scan/pathology layers based on whom the re-
port belongs to with the test report date and symp-
toms as the label information. From this MLN,
it is also possible to extract graphs for an individ-
ual or a select group for different types of analy-
Figure 7: Healthcare MLN.
sis (shown in Figure 7(b) for patient p
1
and his/her
family). This model with the extracted graph(s) al-
lows us to query, search, and analyze to discover
knowledge using all or a subset of layers in various
ways. Few examples are - using collective infor-
mation of an individual and family, a physician can
draw holistic inferences which may not be possible
without a model that represents multi-source, multi-
type data (personalized/customized holistic diag-
nosis/inference), find group(S) of people for a spe-
cific demographics who had lung problems and other
co-morbidity (e.g. diabetes) and contracted Covid
(aggregate analysis using homogeneous and het-
erogeneous community detection on multiple lay-
ers), people who did not have any history of lung
issues but contracted Covid (mining on a subset of
layers using Boolean NOT operation).
8 CONCLUSIONS
In this position paper, we argue for MLNs as a vi-
able alternative for big data analytics. We have dis-
cussed the versatility of MLN models and their ability
to model diverse data, the recent work on MLN model
generation using the EER approach, and efficient
MLN algorithm development for analysis. Based on
MLN work in the literature, we have argued for their
use for modeling and analyzing complex data sets in-
cluding images, videos, and other data types (e.g.,
natural language). There is an ongoing effort to ap-
ply MLNs for information fusion/integration as well.
ACKNOWLEDGMENTS
This work was supported by NSF awards #1955798
and #1916084.
REFERENCES
(2014). Gephi - The Open Graph Viz Platform . http://
gephi.org/.
Multilayer Networks: For Modeling and Analysis of Big Data
389

Aharon, N., Orfaig, R., and Bobrovsky, B. (2022). Bot-sort:
Robust associations multi-pedestrian tracking. CoRR,
abs/2206.14651.
Angles, R. (2018). The property graph database model. In
AMW.
Atrey, P. K., Hossain, M. A., El Saddik, A., and Kankan-
halli, M. S. (2010). Multimodal fusion for multimedia
analysis: a survey. Multimedia Systems, 16(6):345–
379.
Barbieri, D. F., Braga, D., Ceri, S., VALLE, E. D., and
Grossniklaus, M. (2010). C-sparql: a continuous
query language for rdf data streams. International
Journal of Semantic Computing, 4(01):3–25.
Berenstein, A., Magarinos, M. P., Chernomoretz, A., and
Aguero, F. (2016). A multilayer network approach
for guiding drug repositioning in neglected diseases.
PLOS.
Billah, H. and Chakravarthy, S. (2024). Video situation
monitoring using continuous queries. In DEXA,2024,
volume 14911 of LNCS, pages 125–141. Springer.
Billah, H., Santra, A., and Chakravarthy, S. (2024). Lever-
aging video situation monitoring in assisted living en-
vironment. In PETRA, 2024, pages 307–315. ACM.
Boden, B., G
¨
unnemann, S., Hoffmann, H., and Seidl, T.
(2012). Mining coherent subgraphs in multi-layer
graphs with edge labels. KDD ’12, pages 1258–1266.
Bollacker, K., Evans, C., Paritosh, P., Sturge, T., and Taylor,
J. (2008). Freebase: a collaboratively created graph
database for structuring human knowledge. SIGMOD
’08, pages 1247–1250, New York, NY, USA. ACM.
Chen, P. P.-S. (1976). The entity-relationship
model—toward a unified view of data. ACM
transactions on database systems (TODS), 1(1):9–36.
Das, S., Santra, A., Bodra, J., and Chakravarthy, S. (2020).
Query processing on large graphs: Approaches to
scalability and response time trade offs. Data Knowl.
Eng., 126:101736.
De Domenico, M., Sol
´
e-Ribalta, A., G
´
omez, S., and Are-
nas, A. (2014). Navigability of interconnected net-
works under random failures. Proc. of Ntl. Acad. of
Sciences.
Domenico, M. D., Nicosia, V., Arenas, A., and Latora, V.
(2014). Layer aggregation and reducibility of multi-
layer interconnected networks. CoRR, abs/1405.0425.
Holder, L. B., Cook, D. J., and Djoko, S. (1994). Substuc-
ture Discovery in the SUBDUE System. In Knowl-
edge Discovery and Data Mining, pages 169–180.
Ji, J., Krishna, R., Fei-Fei, L., and Niebles, J. C. (2020).
Action genome: Actions as compositions of spatio-
temporal scene graphs. In CVPR, pages 10236–10247.
Kim, J. and Lee, J. (2015). Community detection in multi-
layer graphs: A survey. SIGMOD Record, 44(3):37–
48.
Kivel
¨
a, M., Arenas, A., Barthelemy, M., Gleeson, J. P.,
Moreno, Y., and Porter, M. A. (2013). Multilayer net-
works. CoRR, abs/1309.7233.
Komar, K. S., Santra, A., Bhowmick, S., and Chakravarthy,
S. (2020). Eer→mln: EER approach for modeling,
mapping, and analyzing complex data using multi-
layer networks (mlns). In ER 2020, pages 555–572.
Magnani, M., Hanteer, O., Interdonato, R., Rossi, L., and
Tagarelli, A. (2021). Community detection in multi-
plex networks. ACM CS., 54(3):48:1–48:35.
Melamed, D. (2014). Community structures in bipartite
networks: A dual-projection approach. PloS one,
9(5):e97823.
Otter, D. W., Medina, J. R., and Kalita, J. K. (2020). A sur-
vey of the usages of deep learning for natural language
processing. TNNLS, 32(2):604–624.
Ou, Y., Mi, L., and Chen, Z. (2022). Object-Relation Rea-
soning Graph for Action Recognition. In CVPR, pages
20133–20142.
Padmanabhan, S. and Chakravarthy, S. (2009). HDB-
Subdue: A Scalable Approach to Graph Mining. In
DaWaK, pages 325–338.
Pavel, H. R., Roy, A., Santra, A., and Chakravarthy, S.
(2023). Closeness centrality detection in homoge-
neous multilayer networks. In IC3K 2023, KDIR.
Roy-Hubara, N., Rokach, L., Shapira, B., and Shoval, P.
(2017). Modeling graph database schema. IT Profes-
sional, 19(6):34–43.
Samant, K., Memeti, E., Santra, A., Karim, E., and
Chakravarthy, S. (2021). Cowiz: Interactive covid-19
visualization based on multilayer network analysis. In
ICDE 2021, pages 2665–2668. IEEE.
Santra, A., Bhowmick, S., and Chakravarthy, S. (2017). Ef-
ficient community re-creation in multilayer networks
using boolean operations. In ICCS 2017, pages 58–67.
Santra, A., Komar, K., Bhowmick, S., and Chakravarthy,
S. (2022). From base data to knowledge discovery–a
life cycle approach–using multilayer networks. DKE,
141:102058.
Shi, C., Li, Y., Zhang, J., Sun, Y., and Philip, S. Y. (2017).
A survey of heterogeneous information network anal-
ysis. IEEE Trans. Knowl. Data Eng., 29(1):17–37.
Sol
´
e-Ribalta, A., De Domenico, M., G
´
omez, S., and Are-
nas, A. (2014). Centrality rankings in multiplex net-
works. In Procds. of 2014 ACM conf. on Web science,
pages 149–155. ACM.
Sun, Y. and Han, J. (2013). Mining heterogeneous informa-
tion networks: a structural analysis approach. ACM
SIGKDD Explorations Newsletter, 14(2):20–28.
Wang, A., Chen, H., Liu, L., Chen, K., Lin, Z., Han, J.,
and Ding, G. (2024). Yolov10: Real-time end-to-end
object detection. CoRR, abs/2405.14458.
Wang, J., Sun, K., Cheng, T., Jiang, B., Deng, C., Zhao,
Y., Liu, D., Mu, Y., Tan, M., Wang, X., et al. (2020).
Deep high-resolution representation learning for vi-
sual recognition. PAMI, 43(10):3349–3364.
Yadav, P., Salwala, D., Das, D. P., and Curry, E.
(2020). Knowledge Graph Driven Approach to Rep-
resent Video Streams for Spatiotemporal Event Pat-
tern Matching in Complex Event Processing. IJSC,
14(03):423–455.
Yan, X. and Han, J. (2002). gSpan: Graph-Based Substruc-
ture Pattern Mining. In IEEE International Confer-
ence on Data Mining, pages 721–724.
Zhan, Q., Zhang, J., Wang, S., Philip, S. Y., and Xie,
J. (2015). Influence maximization across partially
aligned heterogenous social networks. In PAKDD (1),
pages 58–69.
Zhang, E., Daum, M., He, D., Haynes, B., Krishna, R.,
and Balazinska, M. (2023). Equi-vocal: Synthesizing
queries for compositional video events from limited
user interactions. VLDB, 16(11):2714–2727.
KDIR 2024 - 16th International Conference on Knowledge Discovery and Information Retrieval
390