Visualization Techniques for Network Analysis and Link Analysis
Algorithms
Ying Zhao
1
, Ralucca Gera
1
, Quinn Halpin
2
and Jesse Zhou
3
1
Naval Postgraduate School, Monterey, CA, U.S.A.
2
Cornell University, Ithaca, NY, U.S.A.
3
JZ Tech Consulting, San Francisco, CA, U.S.A.
Keywords:
Visualization, Data-Driven Documents (D3), Network Analysis, Lexical Link Analysis (LLA), Smart Data,
Automatic Dependent Surveillance-Broadcast, ADS-B.
Abstract:
Military applications require big distributed, disparate, multi-sourced and real-time data that have extremely
high rates, high volumes, and diverse types. Warfighters need deep models including big data analytics, net-
work analysis, link analysis, deep learning, machine learning, and artificial intelligence to transform big data
into smart data. Explainable deep models will play a more essential role for future warfighters to understand,
interpret, and therefore appropriately trust, and effectively manage an emerging generation of artificially in-
telligent machine partners when facing complex threats. In this paper, we show how visualization is used in
two typical deep models with two use cases: network analysis, which addresses how to display and present
big data both in the exploratory and discovery process, and link analysis, which addresses how to display
and present the smart data generated from these processes. By using various visualization tools such as D3,
Tableau, and lexical link analysis, we derive useful information from discovering big networks to discovering
big data patterns and anomalies. These visualizations become intepretable and explainable deep models that
can be readily used by warfighters and decision makers to achieve the sense making and decision making
superiority.
1 INTRODUCTION
The US Department of Defense (DoD) faces chal-
lenges that demand more deep models to produce in-
telligent, autonomous, and symbiotic systems to sup-
port situation awareness and decision making supe-
riority. Military applications require big distributed,
disparate, multi-sourced, and real-time data that have
extremely high rates, high volumes and diverse types.
Warfighters need deep models including big data an-
alytics, network analysis, link analysis, deep learn-
ing, machine learning, and artificial intelligence to
transfer big data into smart data. Warfighters then
can apply the insight and knowledge generated from
big data for decision making and actions. Explainable
deep models will play a more essential role for future
warfighters to understand, interpret, and therefore ap-
propriately trust, and effectively manage an emerg-
ing generation of artificially intelligent machine part-
ners when facing complex threats. Researchers need
consider two requirements for understandable, inter-
pretable, and explainable deep models for warfighters
and decision makers: The first requirement is to show
the process of discovering knowledge, exploring in-
sight from big data and building actionable deep mod-
els. The second requirement is to comprehend the
resultant smart data and deep models. Visualization
provides one of the important components for the two
needs. There are two the research questions to address
in this paper as follows:
1. How to display and present big data, both in the
exploratory and discovery process?
2. How to display and present the smart data gener-
ated from these processes?
We show how visualization is used in two typical
deep models with two use cases: network analysis,
which addresses the first research question and link
analysis, which addresses the second research ques-
tion. They both have the characteristics of discovering
and exploring new and high-value information where
warfighters lack useful information. The process be-
longs to the new frontiers of deep analytics with the
potentials to handle so-called “unknown unknowns”
Zhao, Y., Gera, R., Halpin, Q. and Zhou, J.
Visualization Techniques for Network Analysis and Link Analysis Algorithms.
DOI: 10.5220/0008377805610568
In Proceedings of the 11th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2019), pages 561-568
ISBN: 978-989-758-382-7
Copyright
c
2019 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
561
scenarios: We do not know if there are any unknowns
in a battlespace.
Many commercially-available data manipulation
and visualization tools exist today. Typical spread-
sheet tools (e.g., scatter and line plots, bar and pie
charts, and bubble and radar charts) are widely used
to visualize statistical characteristics to support data-
driven reasoning and decision making. Engineers use
MATLAB, Octave, and Python libraries to display nu-
meric data and analysis results. Developers use Data-
driven Documents (D3) and Javascript to manipu-
late document object models (DOM) within browsers
and generate dynamic and interactive visualizations.
Business users use Tableau to generate insights from
relational databases and data cubes.
In this paper, we focus on several types of visual-
ization and how they are useful for two deep models
such as network analysis and link analysis. Types of
visualization considered in this paper include statis-
tical, topical, network, temporal and geospatial. The
data types considered include unstructured, structured
and network data.
2 NETWORK ANALYSIS
Visualizing data complements the analysis process
and enables understanding of the “why” behind the
“what” is observed (CED3, 2018).
Inferring the structure of an unknown network is
of interest to researchers in the government/military,
academia and industry. Generally, the ground truth
of a network is not known because it is extremely
large or because complete information about it is not
available. Thus, researchers make decisions based on
the inferred network, whose information is still in-
complete, but which acts as the true network. The
degree of incompleteness of information is not gen-
erally known, since the true network is unidentified
and there are no standard techniques to measure their
topological difference. This visualization project al-
lows for the exploration of pre-loaded network exam-
ples, uploaded networks or the creation of new net-
works using the left navigation toolbar
Our visualization presents the output of novel
algorithms previously introduced to infer nodes
and edges in an unknown network (Davis et al.,
2016)(Gera et al., 2017)(Wijegunawardana et al.,
2017)(Chen et al., 2017). Our algorithms utilize sen-
sors that have the capability to detect neighboring
nodes (and their labels) and the edges incident to the
sensor. The algorithms infer the networks through a
combination of (a) random walks, (b) greedily plac-
ing new sensors on a highest degree neighboring node
that has been inferred, (c) greedily placing new sen-
sors on a highest degree neighboring node of the cur-
rent monitor, (d) greedily placing new sensors on
highest undiscovered degree node neighboring an in-
ferred node. The algorithms have a probability based
restarting feature which varies between restarting at
a previously discovered but unmonitored node and
a random teleportation to an unexplored node some-
where in the network. The algorithms stop when there
is an attempt to place a sensor on a node where a sen-
sor already exists.
We say that a monitor on node i detects an edge i j
if i and i j are incident, and i detects the label i j of the
edge (i.e. the monitor discovers the label of the other
end node of i j) (Davis et al., 2016). This then implies
that a monitor on node i detects a node j if there is an
edge i j connecting them. We also allow the monitor
on i to discover the deg j (Davis et al., 2016).
The following screenshots overview the interac-
tive site visualizing the progression in discovering a
network. Figure 1 displays a network before the dis-
covery process.
Figure 1: An example of a network to be discovered.
The first column identifies the search algorithm to
be used, the network to be analyzed, and the statistical
properties of the network as well as the network to be
discovered.
The rest of Figure 1 is divided into four quadrants.
The top left quadrant, shows the network to be discov-
ered, whose nodes and edges turn green and blue, re-
spectively, as the network is being lit/discovered. The
bottom left quadrant shows a temporal progression of
the network as it is discovered, both for nodes and
edges, color-coded with blue and green to match the
first quadrant. Notice that the x-axis doesn’t go be-
yond 60% of monitored nodes, since by that time ei-
ther the whole network is discovered, or there is no
particular strategy needed for the left over part of the
network. The upper right quadrant, shows the left-
over network that has not yet been discovered; it starts
with the original network, and the discovered part is
being taken away. The bottom right quadrant dis-
plays a network’s heat map, a node being colored dark
KDIR 2019 - 11th International Conference on Knowledge Discovery and Information Retrieval
562
green if 100% of the neighbors have been discovered,
and white if 0% of the neighbors have been discov-
ered, with intermediate percentages represented be-
tween white and green.
This visualization was created with the goal of
supporting decision makers in planning the discovery
of a network, by (1) allowing to visually see what por-
tion of the network has been discovered much like a
map would, (2) what percent of the network has been
identified, measured both for edges and nodes, and
(3) what portion of the network has better coverage
through the heat map. The visualization is particu-
larly useful in testing the patterns and performance of
different algorithms as discussed below.
The patterns of how the algorithms evolve through
networks can be visually seen using the three network
quadrants of Figure 1. The same algorithm can be
run several times, and while it can visually be ob-
served, the data can be exported using the “Network
Exporter” and “Network Statistics” on the very left
panel in Figure 1, that capture both the ground truth
and the discovered networks one step at the time. The
performance of the algorithm can be measured using
the bottom left panel in Figure 1, measuring the “Per-
centage of network discovered” by summarizing sev-
eral runs of an algorithm displaying the average and
confidence intervals (using different fidelity such as
µ ± σ, µ ± 2σ, and µ ± 3σ).
While partial network information is sometimes
insufficient to make decisions, it is sufficient to in-
fluence the process of network discovery. Random
walks have been extensively used to explore networks
of all sizes. However, alternative, better algorithms
to light up a network are useful. We created algo-
rithms that infer networks using a minimal amount
of partial information from sampling. We searched
for a sampling methodology that minimized the sam-
ples needed while maximizing the information each
new node reveals about the network, see (Davis et al.,
2016)(Alonso et al., 2017)(Chen et al., 2017)(Craw-
ford et al., 2016)(Wijegunawardana et al., 2017).
We measure the captured information by the per-
cent of nodes and edges sampled nodes can observe.
Each sampled node (called a monitor) detects its
neighbors, and the edges between the monitor and its
neighbors. We introduced variants of how the algo-
rithms progress through the network based on discov-
ered nodes’ degrees and size of the network (Davis
et al., 2016).
We show the progression of a couple o different
types of networks. The first network is an Erd
´
os
R
´
enyi Random Network (ER) network. In this model
the number of nodes and edges are specified, but the
distribution of the degrees is random. This network
type displays no particular structure being random.
The second network is built from four cliques, a mod-
ular network identifying how the algorithms work dif-
ferently if there is structure present.
2.1 First Case Study: A Random
Network
We first consider an ER random graph with 30 nodes
and 40 edges. Figure 2 displays a temporal snapshot
of the first two quadrants on network discovery using
a random walk as the inference algorithm to light up
the network.
Figure 2: An example of network discovery using a random
walk.
Figure 3 displays a temporal snapshot of the first
two quadrants of the same random graph with 30
nodes and 40 edges, at the same point in time us-
ing the highest degree of the discovered node (Davis
et al., 2016) as the inference algorithm to guide the
placement of the next monitor to light up the network.
One can compare the progression in Figure 3 to Fig-
ure 2 in order to see the strength of the algorithms.
2.2 Second Case Study: A Modular
Network with Four Communities
We consider a network of 40 nodes, partitioned into 4
cliques of 10 nodes each, with more edges appear-
ing based on a probability of 0.2 between cliques.
The communities of the network are identified before
lighting it up. Figure 4 displays a temporal snapshot
of the first two quadrants of the network using a ran-
dom walk as the inference algorithm to light up the
network. Figure 5 displays a snapshot of the first two
quadrants of the network using the highest degree of
Visualization Techniques for Network Analysis and Link Analysis Algorithms
563
Figure 3: An example of network discovery guided by the
highest global degree.
Figure 4: An example of network discovery using a random
walk.
the discovered node (Davis et al., 2016) as the in-
ference algorithm to guide the placement of the next
monitor to light up the network.
The above use cases illustrate how the network vi-
sualization discovery tool gives different methods of
network discovery more meaning as we light up the
network of concern. That knowledge can immedi-
ately lend insight into further decisions based on the
discovered network.
3 LEXICAL LINK ANALYSIS
(LLA)
We use LLA as an example of deep models (Zhao
et al., 2015). In a LLA, we describe the characteristics
Figure 5: An example of network discovery guided by the
highest global degree.
a complex system using a list of attributes or features
with specific vocabularies or lexical terms. For ex-
ample,we can describe a system using word pairs or
bi-grams as lexical terms extracted from text data.
Figure 6: An example of a theme discussed in LLA.
Figure 6 shows an example of a word network dis-
covered from text data using LLA. For a text docu-
ment, network nodes represent words, and network
edges or links represent word pairs or bi-grams be-
tween nodes. A list connected words or word pairs
forms a network with the center word “energy” as
shown in Figure 6. “Clean energy, “renewable en-
ergy” are two bi-gram word pairs examples. The bi-
gram method in LLA extends the use of LLA to struc-
tured data and combination of structured and unstruc-
tured data such as data in the social media
We applied LLA in many use cases to greatly fa-
KDIR 2019 - 11th International Conference on Knowledge Discovery and Information Retrieval
564
cilitate the discovery of high-value information in dif-
ferent application domains. LLA outputs smart data
such as semantic and social networks , patterns such
as rules, associations , themes, and topics . The
themes and topics discovered by LLA are further di-
vided into the popular or authoritative, emerging and
anomalous information categories. Information users
can use authoritative information to discover leader-
ship and archetypes in a social network , use emerging
information to discover high-value information from
crowdsourcing , and use anomalous associations to
identify fraudulent behavior and imposters
The output of LLA includes a file of associations
where if word pairs (for unstructured data) or lexical
features (for structured data) are linked together. To
represent this smart data, we use a variety of visual-
ization methods including D3 detailed below. While,
the D3 templates (Bosack, 2017) creates the initial
foundation, we designed and implemented custom
features and class structure for displaying the smart
data output from LLA.
3.1 Use Case: ADS-B Data Analysis
We use a big data set called Automatic Dependent
Surveillance-Broadcast (ADS-B) in the context of the
Naval Common Tactical Air Picture (CTAP) process
and Combat Identification (CID). The CTAP process
collects, processes, and analyzes data to provide sit-
uational awareness to decision makers. The accurate
CID process enables warfighters to locate and iden-
tify airborne objects as friendly, hostile or neutral.
CID plays an important role in generating the CTAP
in the whole kill chain process (Zhao et al., 2016).
We downloaded four terabytes historical ADS-B data,
sampled every minute, for the whole year (6/2016 to
6/2017) (ADSBexchange.com, 2017). We used LLA
to analyze patterns and anomalies in the ADS-B data
in an effort to improve CTAP and CID.
3.2 Exploratory and Discovery Process
Exploring and gaining insight from big data starts
with visualizing basic statistics, correlations and pat-
terns. In this first step, human analysts use visualiza-
tion tools to report basic statistical facts within a data
set and discover initial patterns and correlations. Hu-
man analysts examine the quality of the data, validate
and observe initial patterns and compare the knowl-
edge data with their existing knowledge. Here we
show examples generated from Tableau and Matlab
to display multi-dimensionable (columns, rows, col-
ors and bubble sizes) aggregated data and measures
(e.g.,average, total). Figure 7 uses a bubble chart
to illustrate the frequencies for each type of aircraft,
with the color showing if an aircraft is military or not.
The size of bubble shows the number of flights for
the aircraft type, e.g., B737 has the biggest bubble.
The yellow bubbles represent military aircraft. Com-
mercial aircraft dominate because there are more and
bigger bubbles in the plot. Figure 7 is generated
using Tableau. Figure 8 shows a MATLAB geospa-
tial visualization of the detailed ADS-B tracks, point-
by-point, colored by average absolute heading change
in a track. The higher measure of the average abso-
lute heading change may indicate the activities of tak-
ing off and landing in an airport area for commercial
flights.
Figure 7: Topical visualization: Bubble charts show mili-
tary flights (yellow) or commercial Flights (blue) vs. air-
craft types. The size of bubble shows the number of flights
in the data for the aircraft type.
Figure 8: Geospatial visualization: Tracks colored by aver-
age absolute heading change. Near the airports: The total
heading changes of flights are larger while passing by flights
have smaller heading changes.
3.3 Force Directed Graph
To address the research question 2, we consider one of
the challenges for LLA is that when anomalous asso-
Visualization Techniques for Network Analysis and Link Analysis Algorithms
565
ciations are considered as a high-value information, it
becomes a “needle in a haystack”. In order to find that
needle, one has to compute all the associations and
then filter accordingly. This can be done in the initial
property setting of the LLA tool to some extent. We
investigated how to pre-compute all the associations
and then filter/select to visualize part of them based
on a user’s requirements.
The link or association outputs from LLA are
force directed graphs where each node represents a
word and each link represents a word association. The
nodes are colored according to the node’s anomalous,
emerging or popular type. LLA outputs the types.
The nodes (or words) are also further clustered into
various themes. Each theme contains a group of word
pairs that are related to each other based on the data.
The D3 force directed graphs visualization doesn’t
show large unfiltered data sets well because the pre-
sentation resides in a web browser. The screen begins
to lag with too many nodes. The nodes also cannot
leave the canvas, so as more nodes are featured they
begin to overlap and block other nodes and lead to a
cluttered and disoriented view of the data.
Visualizing an unfiltered data set containing a
large number of nodes which are very slow to be visu-
alized in the browser. After re-designing the conven-
tional D3 template, we are able to display the same
data set with a visible improvement of performance
with little latency issues in Figure 9. We improved
this visualization by providing tools to filter the data
set according to the users needs. For example, one fil-
ter the output associations (smart data) from LLA ac-
cording to the LLA groups (themes), word, or words
that begin with ’x’, end with ’x’, or include ’x’, where
’x’ can be any string of alphanumerics. Other filter
types include filtering by the strengths of the associa-
tions defined using properties initially set by a user.
Such properties include node degrees,data sources,
association types, and strengths for the output associ-
ations. The filter capabilities are important for the re-
search Question 2 because end users often depend on
the parameter filters to execute queries and hypothe-
ses.
The forces of the strength of associations on the
nodes cluster the nodes automatically, making small
clusters naturally lay themselves out in an appealing
and easy to view manner. The nodes can also be
dragged around and forces can be turned off.
Figure 10 an example of the profiles of aircraft
discovered by LLA. Each aircraft track is represented
as a time series of kinematic measures such as po-
sition (latitude, longitude, and altitude), speed, and
heading. Each attribute such as “average altitude”
is computed using aggregation statistics such as av-
Figure 9: LLA output associations filtered using the filter
conditions defined on the right.
Figure 10: Profiles of aircraft discovered by LLA.
erage or total to the time point for a track. Then such
an aggregated track statistics is discretized into three
bins: less than (lt) the mean (of the statistics) sub-
tracting one standard deviation, between (bt) the mean
subtracting one standard deviation and the mean plus
one standard deviation, and more than (mt) the mean
plus one standard deviation. For example, the word
“avg alt mt 31557.5” means average altitude more
than 31557.5 feet. These word features are filtered
using the visualizer in Figure 9 into five dominant
clusters, representing five profiles of the aircraft fly-
ing characteristics as follows.
Normal flights:
Average altitude more than 31557.5 feet
(avg alt mt 31557.5)
Average absolute altitude change between 0
and 743.5 (alt alt chg abs bt -7.0 743.5)
Total absolute heading change between 0 and
150.1 (tot hdg chg abs bt -211.5 150.1)
Low-flying flights
Total absolute altitude change between
0 and 13504.3 feet (tot alt chg abs bt -
KDIR 2019 - 11th International Conference on Knowledge Discovery and Information Retrieval
566
854.9 13504.3)
Average altitude less than 8349.7
(avg alt lt 8349.7)
Total altitude change between -15.8 feet and
16813.8 (tot alt chg bt -15.8 16813.8, nega-
tive change means going down)
Taking off flights
Average absolute altitude change more than
1494.0 feet(avg alt chg abs mt 1494.0)
Average speed change more than 14.4
(avg spd chg mt 14.4)
Total time in the area between 0 and 1124.9 sec-
onds (tot time bt -8.6 1124.9)
Small heading changing flights
Average heading change between -0.9 and 7.4
(avg hdg chg bt -0.9 7.4)
Total heading change between -13.3 and 149.2
(tot hdg chg bt -13.3 149.2)
Slowing down or landing flights
Average speed change between -12.8 and 0.8
(avg spd chg bt -12.8 0.8)
Total speed change between -166.5 and 1.7
(tot spd chg bt -166.5 1.7)
3.4 Dynamic Time Series
Temporal visualization as shown in Figure 11 used on
a LLA output displays sequential patterns. The time
variable date is shown on the x-axis. The size of a cir-
cle represents the degree of ground truth in a data set.
In the ADS-B data set, if a flight is a military or not
(mil=1 or 0) is the ground truth of interest. The visu-
alization shows if and how kinematic attributes (e.g.,
average altitude/speed and change in altitude/speed)
and other attribute such as country of the origin (cou),
and output types (popular, emerging and anomalous)
from LLA are correlated with the ground truth of in-
terest and how the correlation is distributed over time.
This visualization can be used in the exploratory and
discovery process as well as in the post LLA analy-
sis to address both in research question 1 and 2. For
example, in Figure 11, each node is represented using
the date (x-axis), y-axis, radius and category. The y-
axis can represent any of the numeric variables in the
data, avg alt N as the “average altitude” for a track.
The radius can represent any of the numeric variables,
e.g., mil=1 or 0 in Figure 11. The category can rep-
resent any categorical variable, e.g., cou in Figure 11,
using colors. There are three observations that show
the insights of the behavior of the aircraft as we can
obtain from Figure 11:
Most flights’ average altitudes value between 0 to
40,000 feet
Flights can fly over 100,000 feet. These can be
caused by data errors or anomalies.
Military flights with larger radius dominate with
the country of the origin (cou) of the United States
(pink). Some other countries other than the US
(other color than pink, e.g., one as highlighted in
the attribute list next to one of the gray nodes) also
have military flights in the data set, which might
be interesting data to investigate more.
Figure 11: Time series visualization. Radius by mil=1 or 0
(if a track is military or not). Y-axis by average altitude of a
track. Category by cou, i.e., the country of origin.
This visualization shows the same data with the
flexibility of changing the y-axis and radius dynami-
cally depending on the attributes a user chooses. The
user can also hover over a data point (e.g., the point
next to the arrow) to see all of the information dis-
played at once. In addition, because required com-
putations are fairly simplistic, this visualization faces
fewer latency issues compared to that of a Force Di-
rected Diagram.
3.5 Virtual Airways of the ADS-B Data
Figure 12: Zoomed-in look at ADS-B visualization. Each
line represents the trajectories of all planes that take off at
the same specific time and fly in that area over 100 minutes.
The goal of this visualization aims for a user to view
the trajectories of airplanes and then filter out based
on the requirements of different velocity, flight pat-
terns, destinations, and arrivals. This visualization
was created with the D3 projection API . The visu-
alization can display the signal of every plane at any
Visualization Techniques for Network Analysis and Link Analysis Algorithms
567
given moment and animate the planes movement from
minute to minute. It can also display all planes that
take off at the selected time and show where that plane
goes in the scope of all the files loaded into the server.
This visualization is limited to the size of the data set.
To play 10 minutes of flights, ten 5KB files need to
be loaded in and looped through to show all of the
data. Therefore, as more time is displayed, latency in-
creases. Current features of this visualization include
zooming- in and out and panning to different sections
of the world as shown in Figure 12.
4 CONCLUSION
Visualizations help users gain insight from big data.
We showed in this paper various visualizations in or-
der to help users understanding big data as well as to
extend the users’ understanding of smart data through
deep models such as link analysis and network anal-
ysis. The visualizations implemented for LLA and
network analysis vary in complexity and offer some
breadth to the viewers. By using D3, Tableau, and
MATLAB visualizations, we derived useful informa-
tion from discovering big networks to discovering big
data patterns and anomalies.
What are the challenges of future visualization?
Assessing data visualizations includes using heuris-
tic evaluation and user studies. Future work for these
visualizations includes designing and developing vi-
sualization types associated with the nature of deep
models, data types and business problems, and mak-
ing the visualization easy to use for human analysts
both in the pre- and post- analyses of big data. This
should be an ongoing effort to improve understand-
able, intepretable and explainable deep models that
can be readily used by warfighters and decision mak-
ers to achieve superiority.
ACKNOWLEDGEMENTS
Thanks to the Naval Research Program at the Naval
Postgraduate School, the Office of Naval Research
(ONR), and the SBIR contract N00014-07-M-0071
for the research of lexical link analysis and collabora-
tive learning agents at Quantum Intelligence, Inc. The
views and conclusions contained in this document are
those of the authors and should not be interpreted as
representing the official policies, either expressed or
implied of the U.S. Government.
REFERENCES
ADSBexchange.com, L. (2017). Ads-b exchange.
Alonso, L., Crawford, B., Gera, R., House, J., Knuth, T.,
M
´
endez-Berm
´
udez, J., and Miller, R. (2017). Identi-
fying network structure similarity using spectral graph
theory. Applied Network Science.
Bosack, M. (2017). D3 data-driven documents.
CED3 (2018). Center for educational design, development,
and distribution.
Chen, S., Debnath, J., Gera, R., Greunke, B., Sharpe, N.,
and Warnke, S. (2017). Discovering community struc-
ture using network sampling. In 32nd ISCA Interna-
tional Conference on Computers and Their Applica-
tions. Springer.
Crawford, B., Gera, R., House, J., Knuth, T., and Miller,
R. (2016). Graph structure similarity using spectral
graph theory. In International Workshop on Com-
plex Networks and their Applications, pages 209–221.
Springer.
Davis, B., Gera, R., Lazzaro, G., Lim, B. Y., and Rye, E. C.
(2016). The marginal benefit of monitor placement on
networks. In Complex Networks VII, pages 93–104.
Springer.
Gera, R., Juliano, N. R., and Schmitt, K. R. (2017). Opti-
mizing network discovery with clever walks. In Pro-
ceedings of the 2017 IEEE/ACM International Con-
ference on Advances in Social Networks Analysis and
Mining 2017, pages 1217–1224. ACM.
Wijegunawardana, P., Ojha, V., Gera, R., and Soundarajan,
S. (2017). Seeing red: Locating people of interest in
networks. In Workshop on Complex Networks Com-
pleNet, pages 141–150. Springer.
Zhao, Y., Kendall, W. A., and Young, B. W. (2016). Big
data and deep analytics applied to the common tac-
tical air picture (ctap) and combat identification(cid).
In Proceedings of the 8th International Joint Confer-
ence on Knowledge Discovery, Knowledge Engineer-
ing and Knowledge Management, volume 1 of IC3K
2016, pages 443–449.
Zhao, Y., MacKinnon, D., and Gallup, S. (2015). Big data
and deep learning for understanding dod data. In Jour-
nal of Defense Software Engineering, Special Issue:
Data Mining and Metrics.
KDIR 2019 - 11th International Conference on Knowledge Discovery and Information Retrieval
568