Visualization Techniques for Network Analysis and Link Analysis

Algorithms

Ying Zhao

, Ralucca Gera

, Quinn Halpin

and Jesse Zhou

Naval Postgraduate School, Monterey, CA, U.S.A.

Cornell University, Ithaca, NY, U.S.A.

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

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

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.,

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