►
From YouTube: Social Cybersecurity WG: Complex networks, AI and the computational study of terrorism
Description
Presenter: Dr. Gian Maria Campedelli
Institution: University of Trento, Italy
A
It
is
my
pleasure
to
introduce
John
Maria,
who
is
a
postdoctoral
research,
fellow
in
computational
sociology
and
criminology,
at
the
University
of
Toronto
Italy
in
2020.
He
earned
a
PhD
in
criminology
from
Catholic
University
in
Milan.
If
you
from
2016
to
19,
he
worked
as
a
researcher
at
transcribe,
the
joint
Research
Center
on
transnational
crime
of
Catholic
university
university
of
benonia
and
University
of
Perugia
in
2018.
He
was
also
a
visiting
research
scholar
in
the
school
of
computer
science
at
Carnegie,
Mellon
University.
A
His
research
addresses
the
development
and
application
of
computational
methods,
especially
machine
learning
and
complex
networks,
to
the
study
of
criminal
and
social
phenomena,
with
a
specific
focus
on
organized
crime,
violence
and
terrorism.
So
John
over
to
you-
and
you
will
find
a
lot
of
interesting
folks
in
the
team
today,
just
to
listen
to
your
talk
and
I
will
also
request
everyone
to
please
stay
muted.
If
you
have
a
question,
please
use
the
chat
feature
and
I
will
be
moderating
the
chat.
A
A
B
B
So
is
going
to
be
the
title:
complex,
natural,
artificial
intelligence
and
the
computational
study
of
terrorism.
A
lot
of
topics
I'll
try
to
keep
us,
keep
it
as
brief
as
possible.
I'm,
not
very
good
at
timing.
So
please
I'm
in
stop
me
or
give
me
heads
up
if
if
things
are
are
running
out
of
time,
so
who
am
I?
This
is
a
router
philosophical
question.
B
So
the
talk
outline
will
be
structuring
four
different
kind
of
sections.
First,
one
will
be
a
very
short
overview
on
computational,
modeling
terrorism,
research
and
then
I'll
present,
two
studies
that
I
co-authored
that
are
actually
on
this
on
this
field
and
then
I'll
try
to
wrap
it
up
with
some
concluding
remarks.
B
Think
now
he
moved
from
Carnegie
Mellon
to
be
at
the
West
Point
Academy
and
then
Dr
mahila
bartolovic,
recently
awarded
a
PhD
in
computer
science
at
Carnegie
Mellon.
Now
working
as
a
software
engineer
in
the
startup
that
Professor
Kaufman
Carly
found
it
so
very
short
review
on
computational
modeling
terrorism,
research.
B
Since
the
70s
there
have
been
appeals
about
the
state
of
Empirical
research
and
terrorism
and
many
Scholars
over
time
try
to
highlighted
various
methodological
issues
that
terrorism,
research,
empirical
science,
terrorism,
research
had.
These
are
mainly
four.
So
we
have
a
lack
of
good,
empirically,
grounded
research,
so
mostly
literature,
reviews
and
opinion
pieces
theoretical
pieces
without
any
kind
of
empirical
backup,
the
other
Reliance
on
secondary
sources,
which
made
it
very
difficult
to
innovate.
B
So
in
2014,
Mark
sageman
talked
about
this
problem
of
stagnation
of
terrorism.
Research,
especially
in
light
of
the
massive
fundings
that
after
9
11
were
allocated
to
the
field
that
actually
were,
was
not
able
to
fulfill
the
many
problems
and
and
respond
to
many
open
questions
about
terrorism
that
are
still
open.
B
A
newest
review,
published
in
2020
by
Sherman,
actually
was
a
little
bit
more
optimistic
and
he
pointed
out
that
we
are
moving
towards
the
the
solution
of
some
of
the
long-standing
issues,
although
some
are
still
in
place
to
date,
economics
and
political
science.
Among
the
many
fields
that
really
deal
with
terrorism
are
the
ones
with
the
highest
meteorological
standard,
which
is
not
not
surprised,
is
kind
of
common
for
many
social
phenomena,
social
problems
and
in
the
last
10
years
we
have.
B
You
know
two
areas
in
quantitative,
computational,
social
research
that
fostered
a
lot
of
research.
A
lot
of
interest
in
terrorism.
Research
is
to
do
our
Network
science
and
machine
learning
and
predictive
modeling.
So
the
memorable
items
of
terrorist
networks.
Why
do
I
call
it
invariable
lightness?
Because
you
know
we
have
this
spike
in
interest
in
in
terrorism?
From
the
point
of
view,
social
network
analysis,
a
lot
of
studies,
however,
most
applications
will
rely
on
affiliation
networks
or
Alliance
and
rivalries
networks,
which
are
mostly
characterized
by
cross-sectionality
and
blur
temporal
boundaries.
B
However,
after
you
know,
a
lot
of
works
that
came
out
came
out
as
vibrant
debate.
That
models
are
not
able
to
meaningfully
forecast,
for
example,
violence
eruptions.
This
was
a
kind
of
a
critic
that
came
out
from
the
conflict
research
agenda
that
is
really
relevant
also
to
terrorism
research
to
date,
so
the
machine
learning
hybrid,
more
excitement
at
actual
results
possible
causes
the
inability
to
capture
complex
interdependencies
between
events,
the
lack
of
cause
and
knowledge
and
theoretical
reasoning.
B
So
this
talk
in
the
spirit
of
what
I've
just
said.
The
stock
will
try
to
present
to
works
that
hopefully
goes
in
the
direction
of
tackling
some
of
the
problems
that
I
mentioned.
So
the
first
one
is
learning
future
terrorist
Target
through
temporal
metagraphs
games
are
forecasting
future
targets,
leveraging
the
ability
of
graphs
to
capture
retrooperational
dependencies
across
events
and
then
multimodal
metrics
reveal
patterns
of
operational.
Similarity
of
Service
Groups,
basically
present
a
graph
representational
learning
approach
to
detect
behavioral,
similar
ID
between
groups
at
a
global
level.
B
So
here's
the
connection
between
the
content
of
this
talk
and
title,
which
is
the
integration
of
complex
natural,
artificial
intelligence
for
the
study
of
terrorism.
So
the
first
study
we
have
three
issues
with
the
accident:
literature
on
on
terrorism,
operational
information
about
terrorists
or
studies
that
overly
aggregate
temporal
levels
so
like
years
or
months
or
quarters,
which
makes
it
very
difficult
to
have
an
utility
at
a
policy
level.
B
Micro
level
forecasting
only
focusing
on
events
and
lethality
to
discriminate
between
events
by
and
this
overlooks,
the
other
originality
of
attack
characteristics
and
then
computational
attempts
mostly
ignore
Theory.
Blinded
the
letter
data
speak
philosophy,
so
the
aims
of
the
work
are
to
develop
a
forecasting
modeling
framework
to
predict
Target
at
higher
risk
of
being
hit
by
operating,
A,
fine
grain
and
temporal
level,
exploiting
Riches
of
data
in
terms
of
event,
characteristics
and
building
the
power
Theory.
B
So
we
can
think
of
groups
as
a
single
entity
that
makes
decision
and
the
Strategic
decision
making
is
limited
by
several
constraints
that
influence
the
type
of
attacks
that
a
group
or
an
actual
plot
means
that
there
are
some
characteristics
and
factors,
some
variables
that
impact
and
influence
the
way
in
which
this
events
will
occur
and
the
kind
of
nature
of
these
events
also
terrorism,
does
not
occur
at
random.
We
know
that
it
follows
specific
patterns
like
self-exidability,
self-propagation
and
spatial
concentration.
B
C
B
Is
the
the
main,
like
the
Supreme
question,
that
we
try
to
Target
and
the
second
question
was
okay
but
Dan?
How
can
we
integrate
if
this
is
true?
And
if
we
want
to
investigate
it,
how
can
we
integrate
the
two
fundamental
layers,
the
temporal
dependence
and
the
operational
dependence
of
strategic
decision
making
this
graph
kind
of
summarize
it?
We
think
that
the
solution
is
temporal
mathographs,
so
temporal
mathographs
are
able
to
capture
connections
over
time
between
events,
but
also
across
event
characteristics.
B
So
we
move
from
IID
a
research
design
that
are
very
common
in
which
events
are
actually
considered
as
independent
of
one
another.
We
move
to
a
more
kind
of
a
sophisticated
way,
which
is
thinking
about
temporal
dependence,
so
one
event
will
impact
the
probability
that
another
event
will
occur
and
we
go
to
the
temporal
methograph
approach.
Basically,
we
know
there's
a
temporal
dependence,
but
we
also
want
to
investigate
whether
there's
a
value
in
understanding
the
and
their
dependencies
between
event
characteristics.
B
So
we
focus
on
two
countries,
Afghanistan
and
Iraq
for
the
sake
of
numerology
of
events,
but
also
because
they're
quite
different
countries,
even
though
they
are
high
frequency
countries
where
terrorism
was
was
very
prevalent.
But
if
we
see
the
data
down
here,
we
see
that
we
have
different
temporal
distributions.
We
have
different
quantitative
considerations
to
make
about
these
two
countries.
B
B
This
was
were
two
interesting,
let's
say
settings
to
test
our
our
framework,
so
the
analytical
pipeline
here
I
try
to
summarize
it,
but
we
basically
focus
on
each
attacks,
deploy
tactics,
utilize
weapons
and
attack
targets,
as
distinct
Dimensions
tactics
in
the
global
terrorism
database
are
originally
labeled
as
attack
type,
and
by
focusing
on
this
free
Dimension.
We
we
derive
multivariate
two-day
based
time
series
from
2001
to
2018,
in
which
each
time
serious
observation
Maps
the
centrality
of
a
certain
feature
in
its
dimension
in
a
given
unit.
B
So
basically,
the
centrality
of
that
particular
tactics
or
that
particular
weapon,
or
that
particular
Target.
It
is
actually
capturing
how
popular
how
used,
how
prevalent
that
particular
tactic,
weapon
or
Target
was
in
that
particular
time
frame,
which
is
again
a
two-day
based
unit,
and
then
we
forecast
the
most
Central
future
targets,
trying
to
learn
the
patterns
of
Association
in
previous
time
time,
steps
between
tactics,
weapons
and
targets.
B
So
we
try
to
infer
what
will
be
the
next
Target
by
using
information
on
past
attacks
and
on
this
this
passage
reality
values
that
are
a
byproduct
of
the
network.
Representation
of
connections
between
the
different
event
characteristics
and
we
also
test
with
different
input
weights.
The
algorithms
are
tested
with
different
input
weights,
because
we
wanted
to
know
how
bad
how?
How
long
should
we
go
back
in
the
history
of
the
attacks
to
try
to
optimize
predictive
accuracy?
B
The
longer
we
go
and
it
means
that
the
more
stable
the
activity
of
a
group
of
of
a
certain
terrorist
activity
is
the
shorter.
The
time
frame
the
highest
is
the
frequency
with
which
the
actors
that
are
active
in
that
particular
context
are
changing
their
actions
and
we
experiment
with
different
machine
learning
and
deep
learning
approaches.
But
we
also
compare
between
metagraphs
time
series.
B
So
this
idea
that
by
combining
the
temporal
and
event
characteristic
event
characteristics,
networks
and
resulting
graphs,
we
can
gain
more
knowledge
with
shallow
time
series
which
just
used
it
count
of
features
in
each
time
unit.
And
we
wanted
to
understand
whether
you
know
using
our
our
approach
actually
allowed
us
to
to
optimize
our
predicted
accuracy.
B
So
this
is
just
a
graphical
representation
of
how
we
started
the
the
whole
project
by
representing
starting
from
our
tensor
representation,
in
which
we
have
our
days
and
our
matrices
with
our
days
and
with
our
event,
characteristics
and
the
count
of
each
characteristics
for
each
day
which
are
divided
by
Dimensions.
So
we
have
targets,
we
have
talk
takes
and
we
have
weapons
and
from
that
we
derive
our
our
networks.
And
then
we
calculate
our
centralities
as
a
byproduct
of
prevalence
for
each
of
the
event
features
that
we
wanted
to
study.
B
We
test
a
different
algorithmic,
architectures,
so
Baseline.
Basically,
we
wanted
to
forecast
centrality.
Is
that
t
plus
one
are
the
ground
Truth
at
time?
T?
So
basically,
it's
assume
that
there's
no
change
while
was
most
Central
in
the
past.
It's
going
to
be
Central
also
in
the
future,
and
that
feed
forward
neural
network,
simple
class
of
neural
net,
no
design
for
sequence
data.
B
We
wanted
to
test
whether
we
have
a
temporal
structure
and
then
four
different
algorithms
that
actually
bear
a
temporal
sequence
structure,
so
lstm
CNN
with
a
a
1D,
convolutional
filter
and
then
bi-directional
STM.
So
basically
it's
an
expansion
of
the
LCM.
We
sequin.
We
learn
the
sequence
forward
and
backwards
and
then
a
combination
of
CNN
and
ostm,
so
we
stacked
together
a
1D
convolutional
layer,
a
dense
one
and
then
on
lstm
and
another
dance
for
performance
evaluation.
Basically,
the
centralities
are
and
continues
values
so
from
zero
to
one.
B
We
normalize
them,
but
we're
not
really
interested,
and
so
the
algorithms
actually
learn
the
centrality
value,
but
we're
not
really
interested
in
understanding
how
well
they
were
predicting
the
actual,
continuous
value.
So
we
transformed
the
problem
from
a
regression
one
to
a
ranking
one.
So
basically,
we
developed
two
different
measures
of
accuracy.
Element-Wise
accuracy
which
basically
tells
us
is
the
model
able
to
forecast
at
least
one
of
the
two
most
most
Central
round
truth
targets
for
each
time
unit.
It's
some
it's
it's
a
little
trivial.
B
So
we
wanted
to
know
whether
among
the
two
most
Central,
the
algorithm
is
able
to
pick
at
least
one,
and
then
we
have
cell-wise
accuracy.
So
we
want
to
understand
whether
the
model
was
able
to
correctly
forecast
the
whole
two
item
set.
So
we
have
two
round
Truth,
most
more
most
Central
targets.
We
want
to
understand
what
our
algorithm
is
actually
able
to
produce
those
two
correctly.
B
So
the
results,
comparing
architecture
and
models
we
have
Afghanistan
on
the
left
and
Iraq
on
the
on
the
right.
We
have
the
blue
dots
for
the
methograph
algorithms
and
the
lighter
say
cyan
ones
are
the
ones
with
the
shallow
times
here
is
what
emerges
is
that
generally,
the
metagraph
approach
is
far
better
in
predicting
the
next
future
Targets.
This
means
that
our
Network,
a
representation,
actually
allow
us
to
to
have
much
more
information
about.
B
What's
going
on
in
those
two
settings,
we
have
variation
in
terms
of
the
ability
of
the
algorithm
and
the
input
width
for
each
of
the
two
settings.
So
the
first
results
again
is
that
graph,
the
read
time
series
I
performed
shuttle
time
series
and
forecasting
because
we
have
a
reach
representation.
Second,
is
that
bi-directional
lstm
in
both
settings?
B
Let's
say
optimal
means
that
we
have
more
actors
and
that
these
more
actors
are
also
innovating
more
in
the
country.
So
this
makes
it
a
little
tricky
for
the
algorithm
to
to
learn
for
the
batch
configuration
just
to
focus
here.
What
we
see
is
that
on
the
blue,
we
have
the
ground
Truth.
So
the
number
of
time
that
each
of
these
targets
were
on
on
the
on,
among
the
most
the
two
most
Central
in
time
unit,
and
then
we
have
the
predicted
in
in
green.
B
What
we
see
is
that
the
algorithm
is
very
the
algorithm
are
very,
is
very
good
and
and
picking
up
the
two
most
prevalent
while
it
struggles
a
bit
to
Peak
some
of
the
less
prevalent
ones,
which
is
something
that
we
should
and
we
want
to
work
on.
Of
course,
when
data,
when
this
kind
of
complexity
and
heterogeneity
and
and
the
phenomenon
is
linked
to
the
fact
that
we
don't
have
so
much
data,
especially
because
we're
transforming
this
in
time
series
it's
going
to
be
challenging.
B
Probably
some
data
augmentation
technique
will
help
and
we'll
certainly
think
it
through,
and
if
you
have
any
suggestion
at
the
end,
we
I'll
be
I'll,
be
glad
to
hear
so
conclusions
for
this
first
study.
Temporal
method,
graphs
provide
future
context
compared
to
Showtime
series
in
line
with
theoretical
premises
and
our
assumptions
and
hypotheses,
then
the
relevance
of
model
testing
uses
different
amount
of
data,
so
the
behaviors
show
different
products
in
different
countries
and
context,
so
there's
no
one-size-fits-all
solution
and
then
the
promising
approach
in
context
with
high
frequency
to
terrorism.
B
However
limitation,
we
lack
a
distinction
of
different
groups
and
answers
operations
in
this
different.
In
the
same
country,
so
the
Iraq
scenario
is
problematic
for
that
precise
region,
and
then
we
don't
hang
Badge
of
spatial
information
which
probably
might
help
us
getting
more
information
and
maybe
refine
our
prediction
and
the
Kirin
injury
system
does
not
capture
rare
events,
so
911
and
similar
events
gets
lost.
We
have
still
the
to
figure
out
how
to
solid
Black
Swan
problem
so
he's
a
reference.
It
was
a
study
published
in
2021
in
nature
scientific
reports.
B
It's
open
access.
So
if
you
want
to
look
it
up,
it's
it's
it's
it's
there.
Second
study.
Multimodal
networks
reveal
patterns
of
operational
similarity
of
service
of
terrorist
groups.
So,
basically
again,
the
idea
is
to
try
to
combine
machine
learning
or
graph
learning,
with
Network
science
and
complex
metrics,
to
try
to
reveal
something
about
how
terrorism
work.
This
is
not
a
predictive
study,
it's
more
about
kind
of
a
descriptive
inferential
one.
B
Surprisingly
also,
there's
not
a
we
weren't
able
to
find
any
comparative
account
of
heterogeneity
among
terrorist
groups
and
serious
actors
in
a
global
scenario,
which
means
that
we
lack
understanding
of
how
singular
groups
are,
how
different
they
are,
how
they
operate.
Do
they
innovate?
Do
they
change
over
time,
so
we
use
again
the
global
terrorism
database
and
we
consider
terrorist
actors
that
have
plotted
at
least
50
attacks
at
the
global
level,
from
1997
to
2018,
accounting
for
a
total
of
105
groups
and
more
than
42
000
events.
B
And
again
we
focus
on
tactics,
targets
and
weapons.
Here
we
have
a
visual
representation.
First
of
all,
the
number
of
groups
and
their
number
of
attacks.
We
see
that
most
groups
are
actually
around
a
50
to
100
attacks
and
we
have
a
a
small
number
of
outliers,
which
are
mostly
aslamist
actors
that
are
able
to
plug
were
able
to
plot
thousands
and
thousands
of
attacks
in
this
time
period.
We
have
also
the
data
about
years
active,
which
we
see
that
it's
pretty
pretty
pretty
reach
as
a
representation.
B
We
have
groups
that
have
been
active
with
high
frequency,
but
just
for
a
little
amount
of
time,
and
then
groups
that
have
been
active
for
even
more
than
20
years,
for
so
for
all
the
the
time
period
under
consideration,
and
we
also
have
the
disaggregation
of
groups
by
their
ideologies,
see
that
Islam
is
not
angiotism
is
the
most
frequent
one
in
general,
so
we
have
atominationalist
groups
as
a
second
second,
most
frequent,
so
computational
methodology.
What
we
did
here
for
each
year.
B
We
took
this
multi-model
network
framework,
in
which,
basically,
we
had
three
different
matrices
again
for
each
year
and
each
Matrix
was
a
group
by
tactic
Group
by
Target
and
group
by
weapon
Matrix,
where
the
link,
where
weighted
links,
meaning
that
the
weight
was
the
number
of
times
that
that
particular
group
was
separate,
use
that
particular
tactic
or
Target
or
weapon
in
that
particular
year.
So
by
starting
from
this
scenario.
B
Basically,
we
transformed
these
networks,
this
multimodal
framework,
which
is
a
multimodal
bipartite
framework
into
a
multimodal
unit
model
framework
by
using
our
reducible
graph
procedure.
So
basically,
in
the
end,
we
come
up
with
networks
that
connect
groups
and
the
connect
groups
if
they
were
similar
and
use
of
tactics,
targets
and
weapons,
and
that
precise
here
and
then
we
devise
these
multi-view
moderati
clustering
procedure
that
was
first
developed
by
Ian
krugshank
and
it's
a
doctoral
phases
which
allow
us
to
optimize
this
measure
of
modularity
across
different
modes
of
the
networks.
B
And
we
do
this
two
times
so
basically,
the
first
multi-human
or
larger
clustering
computation
is
allow
us
to
filter
out
the
cluster
to
outliers.
So,
basically,
we
filter
out
those
organizations,
then,
in
in
each
of
the
modes
for
each
year,
are
particularly
impressive
in
the
way
and
different
from
the
others
in
the
way
that
they
behaved.
B
So,
for
example,
these
are
the
reducible
graphs
for
97
to
205,
211
and
2018.
For
all
the
free
modes-
and
we
see
that
we
have
outliers,
which
are
basically
isolates,
and
then
we
have
a
core
component
of
groups
that
are
that
seem
very
similar
because
they
are
very
connected.
But
in
the
end,
if
we
look
at
the
nuances
of
the
data
and
the
distribution
of
the
weights
in
each
of
this
networks,
we
see
that
there's
still
something
to
be
discovered
there
and
still
at
originating.
B
That's
why
we
wanted
to
run
a
second
attempt
for
separating
and
refining
the
Clusters.
So
we
detect
some
clusters
in
the
end
again,
the
fact
that
we
use
this
multi
multimodal
base
modularity
allowed
us
to
end
up
with
having
clusters
that
combine
together
the
data
on
tactics,
targets
and
weapons,
and
we
see
that
the
number
of
clashes
doesn't
seem
to
give
us
or
provide
any
relevant
information.
There's
not
a
trend
there.
B
It
seems
pretty
flat,
but
if
we
consider
the
number
of
active
groups
over
time
from
1997
to
2018,
and
then
we
calculate
a
ratio
between
the
Clusters
and
a
number
of
groups,
we
see
that
actually
we
have
a
decreasing
Trend,
meaning
that
probably
this
ratio,
cluster
groups
showing
on
that
one
Trend,
probably
is
indicating
us
as
a
reduction
of
heterogeneity
and
complexity
at
the
global
level.
So
we
add
more
groups,
but
we
have
this
kind
of
similar
number
of
classes
over
time.
B
It
means
that
we
don't
need
more
clusters
to
to
study
and
to
link
the
the
increasing
number
of
groups.
So
the
increasing
number
of
groups
probably
are
groups
that
are
more
similar
to
the
others,
and
we
don't
have
you
know
different
groups
that
are
particularly
editor
genius.
So
we
have
a
general
Sensation
that
probably
heterogeneity
and
complexity
is
reducing
at
the
global
level
over
time
to
try
to
understand
this
fact
better.
B
We
also
look
at
the
stability
of
co-class
drain
so
use
we
use
the
adjusted
random
index
and
the
folks
now
score
to
different
measures
of
how
the
Clusters
were
stable
over
time.
And
what
we
see
is
that
we
have
a
strong
stability
from
2011
to
2018,
meaning
that
group
that
had
a
certain
behavioral
pattern
in
2009
kept
it
quite
fixed
in
the
following
year,
so
they
were
clustered
with
similar
groups
and
those
groups
were
not
changing
their
behavior
in
in
this
time
frame.
B
We
also
have
a
region
of
stability
in
the
Years
2002
2006,
meaning
that
again,
groups
were
generally
clustered
that
were
General
clusters
together
in
the
past,
are
going
to
be
clustered
together
in
the
future,
meaning
that
the
two
are
maintaining
their
same
kind
of
Behavioral
profile.
However,
before
2002
we
have
a
high
variance,
so
this
is
possibly
due
to
new
actors
coming
in
or
radioactive
groups
that
are
changing
their
operations
significantly
over
time
and
the
fact
that
group,
a
and
Group
B,
where
it
costs,
are
together
into
1998.
B
So
our
idea
was
also
to
understand
what
drives
Dental
clustering.
What
what's?
What's?
What
are
the
drivers
that
made
groups
being
clustered
together
in
in
a
year,
because
this
is
a
sort
of
a
way
to
understand
again
what
drives
operational
similarity
and
we
wanted
to
test
some
hypotheses
that
we
had,
because,
mostly
in
the
literature,
we
had
the
separation
between
groups
that
is
mostly
based
on
their
ideology
or
their
geographical
setting.
B
But
our
intuition
was
that
maybe
we,
if
we
look
at
operational
characteristics
and
their
behavioral
profiles,
we
can
see
that
maybe
the
fact
that
two
groups
are
of
the
same
ideology
or
do
not
have
the
same
ideology
doesn't
mean
much,
and
we
can
see
something.
We
can
see
some
similarities
at
a
behavioral
level,
even
when,
when
two
groups
are
fighting
for
two
very
different
reasons,
so
we
use
exponential
round
graph
models
on
the
fully
connected
component,
like
networks,
the
rifle
cluster.
B
So
basically
each
group
it's
going
to
be
connected
to
all
the
other
groups
that
are
in
the
same
cluster.
These
give
rise
to
fully
connected
networks
or
isolates
when
we
have
groups
that
are
just
isolating
their
own
cluster
and
we
run
this
separate
year
by
year.
Models
using
exponential,
random
graph
modeling.
B
So
we
look
at
some
of
feature
weights
like
is
the
amount
of
resources
activity,
a
driver
of
cool
clustering
and
the
answer
ECS,
as
we
see
on
the
top
left
some
of
future
subplot
when
the
when
the
dots
are
are
blue.
It
means
that
the
result
is
significant
and
we
see
that.
Yes,
the
fact
that
two
groups
share
the
same
amount
of
resources
so
we're
active
and,
in
the
same
way,
to
the
set
to
the
same
extent
to
another
that
was
a
driver,
cool
class
drain.
B
The
number
of
knowns
there
were
features
in
your
Regional
networks
is
also
repertoire
is
also
a
driver
of
co-clustering,
and
this
is
a
repertoire
of
diversity.
So
the
fact
that
two
groups
are
very
little
diverse
or
very
high
diverse.
The
fact
that
two
groups
have
this
characteristics
are
are
correlated
with
the
fact
they're
going
to
be
clustered
together.
B
So
if
we
have
one
group
that
it's
not
really
diverse
and
another
one
that
continues
to
differentiate,
Direction
they're
not
going
to
be
clustered
together,
however,
the
country
holds
true,
so
you
know
if
you
are
very,
not
not
very
Innovative
and
not
very
diverse
you're,
going
to
be
clustered
with
similar
groups
and
actors
yeah.
We
also
look
at
most
common
targets,
most
common
tactic
and
most
common
weapons.
B
So
the
fact
that
you
use
the
same
weapon
mostly
or
you
deploy
the
same
tactic
or
you
go
against
the
same.
The
most
common
Target
are
these
drivers
of
of
operational
similarity
overall,
not
really
much.
We
have
some
significant.
We
have
some
years
for
which
this
some
of
these
are
significant,
but
mostly
we
don't
wear
any
statistical
significance
in
understanding
similarity
as
a
byproxy
of
sharing
the
same
most
common
Target
tactic
and
weapon
and
I.
B
So
this
means
that
the
fact
that
two
groups
are
operating
in
the
same
scenario
like
Eastern
Europe
or
western
Africa
or
South
America,
doesn't
mean
that
they
have
higher
chance
to
be
operationally
similar,
and
the
most
interesting
one
to
me
is
that
the
fact
that
ideology
is
not
really
a
driver
operational
similarity.
So
the
fact
that
two
groups
are
sharing
the
same
ideology
doesn't
mean
much
we
might
have.
B
So
the
fact
that,
for
example,
two
group
are
Jedis
groups
doesn't
have
doesn't
doesn't
lead
to
a
higher
chance
of
being
clustered
together
than
one
group.
That
is
a
Christian
one
and
one
group
that
is,
she
had
respawn,
and
this
also
holds
true,
for
example,
for
groups
that
are
left-wing
and
right-wing
groups,
so
we
have
Hunter
kiss
or
communist
groups
that
are
very
similar,
for
example,
to
racist
and
Nazi
Nazi
groups
that
are
active
in
the
time
frame
under
under
analysis.
B
So
the
conclusion
of
the
second
study,
multimodal
graphs
allowed
to
capture
operational
similarity
across
terrorist
groups.
This
again
is
a
good
fact
that
embedding
together
complex
networks,
and
so
the
network
representation
of
behaviors
Beyond
physical
representation
of
connections
between
groups
is
important
to
gain
knowledge
about
how
they
operate.
And
then
we
detect
a
reduction
in
our
genetic
variation,
operational
behaviors
in
the
last
years,
and
a
consequent
consequent
higher
degree
of
clustering
stability.
D
B
Time
and
then
from
the
rgm
results,
we
see
that
the
yearly
amount
of
resources,
activity
and
repertoire
diversity,
Drive
code,
clustering
and
different
directions,
and
also
the
the
combined
measure
of
the
two
is
also
a
predictor.
However,
we
see
that
sharing
the
same
ideology
and
acting
in
the
same
region
they're
not
correlated
with
operational
similarity
and
I.
We
think
that
this
goes
against
most
of
the
public
debate
about
terrorism
that
is
currently
heard
on
media
and
and
even
in
academic
Outlets.
B
B
B
But
what
about
terrorism?
So
terrorism,
research
has
certainly
been
lagging
behind.
The
trend
is
similar
to
the
one
witness
and
crime
research,
though
so
we
see
that
there's
a
an
increasing
interest,
although
with
with
a
slower,
Pace,
probably
and
there's,
certainly
an
increasing
fascinated
fascination
with
computational
methods
sparked
by
especially
by
the
opportunities
offered
by
social
media
data
availability.
B
Areas
of
Scholars
criminologists
have
the
main
knowledge,
but
do
not
have
the
tools
to
really
apply
what
they
know
and
to
create
meaningful
models
for
both
research
and
policy,
and
then
terrorism
research
has
seen
it
as
playground
again.
The
problem
of
one-timers
following
the
hype
for
the
AI
for
social
good
Trend,
that,
after
the
probably
2016
so
after
Isis,
you
know
when
Isis
seemed
to
ended
to
be
a
problem
for
the
Western
World.
B
But
this
one
help
us
having.
You
know
the
ability
to
create
solid
infrastructures
and
solid
research
communities.
Then
data
scarcity
beyond
the
gdd,
which
is
again
the
the
best
source
and
the
one
that
I've
been
working
with
since
the
since
my
PhD
there's
no
easily
accessible,
comprehensive,
reliable
data
set
Beyond
this
one,
which
is
which
is
a
problem.
B
Another
one,
is
that
there's
a
small
integration
between
the
offline
and
the
online
world,
so
Fox
working
with
the
gdd
are
just
working
on
the
offline
work
and
then
folks,
working
with
Twitter
data
just
working
on
the
on
their
online
world,
they're
not
talking
to
each
other.
There's
no
integration
want
to
think
there'd
be
certainly
promises
and
trying
to
look
how,
for
example,
the
offline
were
input,
the
online
word
and
vice
versa,
and
then,
which
is
probably
the
most
important
one,
most
technical
ones.
B
The
fact
that
we
have
a
rare
imbalance
Phenom
here
so
predictive
modeling
with
crime,
relies
on
millions
of
observations
as
millions
of
crime
occur
around
the
world
every
year,
while
recorded
terrorism.
Events
luckily
are
just
around
2000,
probably
more
now
for
gdd
data
from
1997
to
today.
So
we're
talking
about
a
very
different
amount
of
information
that
we
can
rely
upon
for
our
for
our
studies
and
for
our
research.
B
So
thanks
for
your
attention,
I
hope
I
was
good
with
times
so
I
think
we're
around
40
minutes
and
if
you
have
questions
I'll
be
happy
to
to
answer
to
them
or
critiques
or
whatever
we
have
to
discuss.
This
is
my
email
by
the
way.
So,
if
you
have
to
go-
and
you
want
to
drop
me
a
line
afterwards,
my
email
is
open
and
I
also
Tweeter,
so
I'll
be
glad
to
to
you
from
from
from
the
stock
on
those
channels
as
well.
Yes,
so
that
was
it.
A
Thank
you.
Thank
you,
John
very
much,
I'm
sure
this
is
a
very,
very
interesting
talk
and
there
must
be
some
thoughts
and
questions.
Let's
open
the
floor
for
discussion
and
I'm
also
monitoring
the
chat.
So
if
you
have
any
questions,
you
can
also
use
a
chat
feature
but
unmute
yourself
and
go
ahead
with
the
question.
C
Hey
John
question
about
I:
guess
how
you're
calculating
or
how
do
you
derive
similarity
I
think
you
kind
of
you
mentioned
it
some
in
here,
but
you
know,
as
you
go
through,
and
look
at
I
was
trying
to
to
bring
all
this
together.
But
when
you
talk
about
similarity
in
that,
you
know
I
forget
which
slide
it
was.
But
you
talked
about.
You
were
talking
about
co-clustering
between
targets,
tactics,
weapons,
ideology,
yeah.
How
are
you
actually
calculated?
Okay,.
C
No
I
think
it
was
the
one
where
you
were
showing
us
towards
the
end.
What
was
driving
co-clustering,
oh
okay,
yeah,
but
I
was
just
curious.
How
are
you
calculating
simul
similarity
because,
like
this
you're
showing
here,
I,
guess
co-clustering,
but
then
you've
already
prior
to
this
you've
already
drive
that
they
were
either
similar
or
not?
Similar
is
the
network
based
on
similarity.
B
B
So
the
fact
that
two
groups
are
in
a
cluster
are,
by
definition
similar
now
the
fact
that
so
what
we
wanted
to
study
is
okay,
let
us
see
what
are
the
variables
that
we
didn't
use
and
in
the
original
Network
representation
that
are
correlated
and
that
are
drivers
of
this
similarity.
So
at
the
beginning
of
our
framework,
what
we
do
is
we
have
this
modal
networks
for
each
for
each
year
and
for
all
the
free
dimension
and
I
mentioned.
So
we
have
tactics,
targets
and
weapons.
B
We
have
this
group
by
type
of
tactics
Group
by
type
of
weapon
Group
by
type
of
Target
networks
that
are
bipartite
networks,
where
the
where
the
weights
are
actually
the
count
of
the
instances
like
I,
don't
know.
I'm,
Isis
and
I
use
in
2015
I
use
20
bombing,
something
like
that.
So
you're
going
to
be
have
20.,
so
we
have
this
weighted
bipartite
networks
which
are
modal
again,
and
then
we
transform
this
bipartite
networks
using
this
radius
ball
graph,
which
is
some
sort.
B
It's
in
computational
geometries
some
sort
of
a
nearest
neighbor
problem.
So
basically
we
take
these
two
by
this
Viper
type
networks
and
we
transform
them
into
Group
by
group
networks,
and
in
that
case
two
groups
are
going
to
be
connected
together
if
they
share
a
lot
of
the
weights
that
were
originally
placed
in
the
networks.
B
The
most
similar
The
Rose
of
the
the
bipartite
networks
are
for
two
groups,
the
more
likely
that
they're
going
to
be
connected
in
the
in
the
unimodal
network.
Now,
once
we
have
those
networks,
we're
gonna
run
our
clustering
procedure,
because
the
fact
that
two
groups
are
going
to
be
connected-
it's
not
per
se
Insurance
of
similarity,
because
maybe
one
one
one
group
is
connected
to
many
other
groups,
but
the
but
the,
but
the
connections
have
different
weights.
B
So
we
want
to
disentangle
that,
and
we
want
to,
we
wanna,
you
know,
derive
what
are
actually
the
strongest
links
to
this
to
do
other
groups
and
that's
how
basically
we
we
then
create
the
links
we.
We
then
create
the
clusters
by
looking
at
the
links
of
the
unimollo
networks,
so
basically
well,
okay,
yeah.
So
basically
we
have
our
bipartite
networks
here,
which
are
this
ones,
and
then
we
we
convert
them
into
monopertite
and
we're
gonna
have
this
once
here.
These
are
the
modern
protect
networks.
B
We
see
that
our
outliers
are
not
connected
to
anyone.
These
are
the
groups
are
very
outstanding,
but
then
you
have
a
critical
mass
of
reserves
that
seem
to
be
very
connected,
and
you
know
at
the
first
glance
you
might
say
they
are
they're
similar,
but
actually
the
way
of
the
weights
of
this
connections
are
very
different.
So
basically,
the
clustering
procedure
allow
us
to
also
extrapolate
similarity
between
those
critical
masses,
and
then
we
have
these
clusters
that
actually
optimize
across
the
free
different
modes.
B
And
then,
when
we
have
these
clusters,
what
we
do
is:
okay,
now,
let's
try
to
use
variables
that
we
that
we
actually
didn't
use
some
of
them.
We
actually
use
them,
but
but
some
of
them
were
actually
embedded
in
the
procedure,
but
some
others
were
not.
So
we
wanted
to
understand
whether
okay,
we
have
the
sum
of
feature
weights-
probably
it's
going
to
be
correlating,
but
in
which
Direction
and
then
the
same
with
number
of
non-zero
features
in
in
which
direction
as
well.
B
But
then
we
wanted
to
understand
whether
what
we
derived
is
it
actually
associated
with,
for
example,
the
same
ideology
or
the
same
region,
or
we
can
actually
see
that
there's
a
operational
similarity
going
on
operational
similarity
pattern
going
on
also
for
groups
that
are
very
distant,
very
far
away
geographically
or
very
far
away
and
their
ideological
Spectrum.
So.
C
No,
that
that
is
helpful,
yeah.
That's
what
I
was
wondering
about
yeah,
so
I
mean
you
are
taking
multiple
steps
to
do
that
and
I
guess.
When
you
get
those
kind
of
the
monopod,
the
monotype
graphs,
I,
guess
part-time
graphs
are
those
are
you
measuring
any
kind
of
I
guess,
I'm
thinking
something
in
terms
of
like
clustering
coefficient
modularity,
something
where
you
can
kind
of
measure
the
the
intensity
or
the
strength
of
that
similarity?
Oh.
B
Well,
that's
a
very
good
question.
Actually
we
also
measure
clustering
coefficient
degree
of
solitivity
degree,
centrality
and
stuff
like
that.
I,
it's
all
in
the
paper,
I
I
didn't
include
the
plots
here.
I
didn't
include
a
plus
here,
but
we
also
look
at
that
to
see
how,
over
time,
for
example,
the
networks
change.
So
we
do
see
more
clustering
over
time,
which
is
also
another
proxy
of
saying
for
saying
that
probably
is
more
of
engineering
going
on
over
time.
B
D
Yeah
thanks
this
is
great
I,
have
a
question
about
sort
of
about
the
distribution
of
the
data
within
the
GTD
and
how
that's
impacting
both
papers
right.
So
we
know
that
there
are
things
like
I'm
going
to
say:
80
of
all
attacks
fall
into
either
the
they're,
either
bombings
or
shootings,
or
something
like
that
right.
D
I.
Think
to
illustrate
the
point
like
we're
doing
so
much
flattening
of
the
Nuance
of
what's
Happening
from
event
to
event
in
the
source
data
that
I'm
actually
really
curious.
What
happens
when
these
methods
are
run
on
data
that
have
far
more
specificity
about
what
happened
right?
Like
did
you
guys
dig
into
sort
of
the
weapon,
subtypes
the
Target,
subtypes,
or
even
consider
how
this
is
different?
B
B
One
thing
that
I
didn't
mention
before
is
that
we
use
all
the
four
four
possible
weapons,
three
possible
attack
types
and
free
possible
targets
for
each
for
each
observation
in,
in
both
cases,
both
the
predictive
and
you
know,
co-clustering
stuff.
So
we
didn't
stop
at
just
the
primary
one
or
the
First
Column,
which
is
something
that
unfortunately
many
people,
many
people
do
it
in
this
video.
B
B
Now
the
using
the
sub
subtypes
I
tested
in
the
early
days
and
then
I
stopped
that
just
because
of
the
fact
that,
unfortunately,
the
the
number
of
observations
here
are
not
as
much
as
many
as
as
you
know
that
they
they
don't
allow
us
to
go
with
a
feature
space.
That
is
gigantic,
so
we
would
run
easily
into
the
crystal
dimensionality.
So
definitely
that
that
that's
that's,
you
know,
that's
limitation
of
the
study
of
the
first
study
for
the
second
one.
I
see
it
as
a
limitation
and
I.
B
Don't
I,
don't
think
that
there
will
be
a
problem
in
testing
the
approach
using
subtypes,
because
in
the
end
we
are
using
here
a
sample
of
groups
that
are
pretty
dear
genius.
Even
though
we're
talking
about
you
know,
the
majority
of
them
are
as
yetis
or
animation
all
this,
but
they
are
as
yet
as
an
animation
release
that
have
very
different
characteristics.
So
we
have
rigidis
and
animationists
groups
that
operate
in
different
geographical
contexts
and
we
with
different
resources
again.
B
So,
yes,
you
know
we
have
the
usual
usual
suspects
that
generally
use
the
same
bombings,
and
you
know
firearms
and
stuff
like
that.
But
we
also
see
a
lot
of
energy
in
general
for
for
some
groups
that
have
some
particular
ideologies
like
I,
don't
know
an
environmental
animal,
animalistic
groups
or
fall
left
groups
and
stuff
like
that,
and
then
one
way
that
is
embedded
in
the
system
that
prevent
this
kind
of
flattening
so
flattening
off
of
the
various
characteristics
is
the
fact
that
we
use
weights.
B
So
we
we
just
don't
use
the
fact
that
that
group
use
a
firearm
to
to
carry
out
an
attack,
but
we
look
also
at
how
many
times
did
you
use
that
and
by
using
this
multimodal
framework?
We
also
implicitly
look
at
the
connections
like
how
many
times
you
use
a
firearm,
but
how
many
times
is
a
firearm
against
like
a
a
police
officer
or
how
many
times
do
Firearms
against
a
civilian.
B
So
yes,
if
we
look
at
the
single
modes
and
single
variables
alone,
you
know
there
they
might
seem
flat-
and
this
is
probably
true
mostly
for
the
predictive
predictive
paper,
but
still
we
might
have
that
concern,
and
also
in
the
second
one.
But
the
precise
objective
of
the
of
the
of
this
papers
was
to
try
to
combine
and
to
try
to
represent
a
complexity
of
terrorism,
behaviors
beyond
the
single
source
of
information
which
might
be
tactic
Target
and
weapons.
B
You
know
the
the
complex
spectrum
of
combinations
of
behaviors,
the
the
frequency
and
and
the
aboriginality
coming
out
of
that
so
I
agree
with
you
that
the
optimal
model
would
especially
in
the
predict
okay,
predict
predictive
case,
would
would
pick
up
more
a
new
and
said
variables
like
the
subtypes
and
the
sub
targets
and
stuff
like
that.
I
think
we
have
three.
You
have
three
levels
of
targets,
I'm
very
afraid,
and
that
was
the
first
feeling
that
we
had
when
we
tested
that
with
this
numerosity.
B
With
this
number
of
observation
that
would
end
up
being
completely
meaningless.
So
we
figured
okay,
it's
better
to
have
a
model
that
is
a
little
bit
less
rich
but
meaningful,
rather
than
having
a
structure,
a
data
structure
that
is
very
rich,
very
nuanced,
but
that
bring
no
noise
out
at
all.
So
that
was
our
kind
of
a
trade-off
and
the
decision
on
how
to
design
the
the
system,
especially
in
the
Memphis
paper.
I,
don't
know
if
that
answers
the
question,
but.
A
B
A
E
Thank
you,
Dr
Joseph.
This
was,
like
you
know,
an
insightful.
The
findings
are
really
insightful.
I
just
had
like,
like
this
a
question
about
this
slide,
exactly
the
results.
What
drives
cochlea
starting
the
first
two
things
like
I,
wanted
to
ask
you
the
amount
of
resources
activity.
What
did
you
mean
exactly
by
that
the
amount
of
resources
activity
and
also
the
repertoire
diversity?
So
yeah?
Can
you
just
explain
on
these
two
sure.
B
So
when
we
we,
when
we
constructed
this
variable,
we
looked
at
the
original,
bipartite
Networks,
so
again,
a
network
in
which
a
model
Network
in
which
we
have
for
tactics,
weapons
and
targets
we
have
for
each
for
each
year.
We
have
our
groups
and
then
the
tactics,
weapons
targets,
and
we
count
the
number
of
times
that
that
group
particularly
use
Firearms
or
use
bombings
or
use
whatever.