Top 5 Benefits OF Using Muvi Live Paywall For Live Streams
Segmentation - The Shadowy Side of Persona Development
1. Segmenta(on:
The
Shadowy
Side
of
Persona
Development
UPA
2012
David
A.
Siegel
Ph.D.
Dray
&
Associates,
Inc.
Minneapolis,
MN
USA
david.siegel@dray.com
www.dray.com
+1
612
377
1980
u Copyright,
Dray
&
Associates,
Inc.,
2012
Copyright 2012
2. Segmentation
Market
Segmentation
User
Classification
2 Copyright 2012
5. • Colors = dimensions
• Can you align them all?
• The most successful are
those willing to break a
partial alignment and start
from scratch
Goal:
Ø Make
explicit
choices
and
tradeoffs,
whether
working
with
and
exis<ng
segmenta<on,
or
proposing
a
classifica<on
scheme
of
your
own
Themes:
Ø Segmenta<on
as
a
subtype
of
classifica<on
Ø Classifica<on
=
selec<ng,
defining,
priori<zing,
and
combining
dimensions
to
usefully
divide
up
a
mul<-‐dimensional
space
Ø Influenced
by
subjec<ve
choices
and
prone
to
distor<ons,
whether
done
casually
or
through
the
most
high-‐powered
sta<s<cal
analysis
6 Copyright 2012
6. Ø What
makes
a
useful
classifica<on?
Ø Tensions
between
marke<ng
and
UX
segments
Ø The
paradox
of
“precision”
Ø Pros
and
cons
of
• Demographics
• Occupa<onal
Roles
• Psychographics
• Behavior
Ø Tensions
between
marke<ng
and
UX
segments
7 Copyright 2012
15. The dimensions we perceive and identify depend
on
Ø Context of comparison
Ø What have we sampled
Ø What distinctions we perceive or assume to be
relevant
E.g., if our purpose was to evaluate agricultural
products in terms of potential for industrialized
production, we might have classified differently
16 Copyright 2012
16. Informa<on
is
a
difference
that
makes
a
difference.
-‐-‐Gregory
Bateson
Source: http://www.nndb.com/people/169/000100866 /
Segmenta<on
needs
to
point
to
different
ac<ons
that
are
available
to
us,
on
the
basis
of
predicted
differences
in
response
from
different
audiences
or
users.
17 Copyright 2012
17. Different
differences
make
a
difference,
depending
on
what
different
ac<ons
we
are
focusing
on.
18 Copyright 2012
18. Classifica<on
Variable
1
Descriptor
Descriptor
Dimensions
Dimensions
Classifica<on
Ac<on
Ac<on
Variable
2
Implica<ons
Implica<ons
Descriptor
Descriptor
Dimensions
Dimensions
Ac<on
Ac<on
Implica<ons
Implica<ons
Ø Not
necessarily
2
x
2,
or
even
factorial
Ø Choice
of
classifica<on
variables
usually
based
on
what
we
think
makes
cleanest
split,
is
easiest
to
detect,
or
summarizes
the
profile
of
descriptors
Ø But
descriptors
could
be
turned
into
classifiers,
depending
on
what
maers
19 Copyright 2012
19. Paradox
of
Precision:
The
“Zoom
In”
Problem
Ø Zoom in = more detailed, granular
description
• More dimensions
• More distinctions
• More subgroups
Ø Perceived
as
more
precise,
more
convincing
Ø But
(all
things
else
being
equal)
finer
grained
dis<nc<ons
become
more
fuzzy,
boundaries
blur
Ø A
law
of
nature!
20 Copyright 2012
20. Case
in
point:
Let’s
zoom
in
here
Non
Customers
customers
Opportunity
At
Risk
Aachment
21 Copyright 2012
21. Aachment
Anything
we
do
to
improve
the
ra<o
of
people
in
our
sample
that
we
are
interested
in
will
exclude
some
of
them,
and
reduce
our
ability
to
know
how
they
relate
to
the
popula<on
as
a
whole
22 Copyright 2012
23. Case
Example:
Segments
based
on
abtudes
did
differ
in
composi<on.
But….
Seg.
A
Seg.
B
Seg.
C
Seg.
D
…the
groupings
across
segments
were
more
coherent
and
dis<nct
re:
usage
paern
24 Copyright 2012
26. Overall
SESS
w
rapidly
Age
ns
mul<ply,
n
simple
rip<ons
‹#›
27. Overall
SESS
w
rapidly
Age
ns
mul<ply,
n
simple
rip<ons
Ethnicity
‹#›
28. Overall
SESS
w
rapidly
Age
ns
mul<ply,
n
simple
rip<ons
Ethnicity
‹#›
29. Overall
SESS
Orienta<on
to
self-‐service
w
rapidly
Age
ns
mul<ply,
Net
Worth
n
simple
Disposable
Income
rip<ons
Ethnicity
Source
of
influence
Importance
of
Iden<ty
‹#›
30. Dimensions
apply
to
all,
but
are
called
out
only
where
most
dis<nc<ve,
heightening
percep<on
of
difference
Affluent
Others
• Highly
Influenced
by
family?
• Highly
Influenced
Hispanics
• High
SES?
by
family
• Manage
own
finances
on
line?
• Manage
own
?
Others
finances
on
line
31 Copyright 2012
31. Segments
Summarizing
Overall
Difference
in
Profile
on
Mul<ple
Dimensions
Ø Some
dimensions
differen<ate
more
strongly
than
others.
Ø Smaller
differences
should
be
weighted
less
Ø But
ofen
all
the
differences
become
equal
parts
of
the
descrip<on
32 Copyright 212
Copyright 2012
32. Some<mes,
it
may
look
like
we
can
make
precise
dis<nc<ons
based
on
small
differences,
only
because
large
samples
make
them
sta<s<cally
significant.
But
do
those
differences
maer?
Math
scores:
Yes,
the
distribu<ons
are
different
(assuming
large
N).
But
if
you
made
dichotomous
decisions
based
on
gender,
(e.g.,
pubng
girls
in
low
math
group
and
boys
in
high
math
group)
you
could
be
wrong
large
%
of
cases.
Source: http://www.wanoah.co.uk/?p=37
The
larger
the
sample
it
takes
to
find
a
sta<s<cally
significant
difference,
the
less
likely
it
is
to
have
a
prac<cal
significance!
33 Copyright 2012
33. Striving
for
Precision
on
Mul<ple
Dimensions
Two
views,
same
points
in
space
à
34 Copyright 2012
34. Back
to
our
unnaturally
clean
clusters:
35 Copyright 2012
35. Imagine
they
are
really
in
3
dimensions,
but
we
have
only
viewed
them
from
one
angle
(i.e.,
only
focusing
on
2
dimensions)
36 Copyright 2012
36. Now
we
rotate.
Same
points
different
views—clusters
smear
out.
37 Copyright 2012
37. In
these
dimensions,
clusters
are
broader,
have
different
members,
personas
not
as
“well
placed”
to
represent
them.
Op<mizing
groupings
on
some
dimensions,
tends
to
“smear”
them
on
38 Copyright 2012
38. Now
imagine
if
we
had
started
off
with
a
more
realis<c
set
of
clusters
—just
slight
varia<ons
in
density
–
because
on
many
of
the
dimensions
we
care
about,
people
don’t
fall
into
such
discreet
groups.
39 Copyright 2012
39. Is
vanilla
ice
cream
more
like
chocolate
milk
or
banana
yogurt?
?
Needed:
a
way
of
combining
differences
on
mul<ple
dimensions
into
a
judgment
of
overall
similarity
and
difference.
40 Copyright 2012
40. Ø Sta<s<cal
approaches
to
building
clusters
usually
try
to
manage
problem
of
over-‐op<mizing
on
some
dimensions
and
smearing
on
others
Ø Use
“distance”
to
represent
“difference”
Heavy
user
of
very
few
features
Cluster
2
Group
A
=
Heavy
Y
=
overall
users
of
many
usage
(me features
Which
group
is
this
one
most
like?
Group
B
=1
ight
to
medium
Cluster
L
users
of
few
features
X
=
number
of
features
used
41 Copyright 2012
41. But
how
do
we
measure
distance?
You
get
to
choose,
e.g.:
Euclidean:
Hypotenuse
of
difference
on
x
and
difference
on
y
ΔX
+
ΔY:
Sum
of
the
differences
on
each
separate
dimension
Both
make
intui<ve
sense,
but
give
different
results!
The
two
methods
assign
the
point
to
different
clusters.
Euclidean
distance
≈
4.6
ΔX
+
ΔY
distance
≈
6.5
Also,
the
dimensions
should
be
weighted
Centroid
of
Group
A
differently
based
on:
Ø Are
they
scaled
the
same?
Ø Are
they
measured
Euclidean
distance
≈
5.94
ΔX
+
Δ
Y
distance
≈
6
equally
reliably?
Ø Are
they
equally
good
predictors
of
Centroid
of
Group
B
something
we
care
about?
42 Copyright 2012
42. Exaggera<ng
Dis<nc<veness
These
look
dis<nct,
but
most
of
Cluster
2embers
have
a
lot
in
their
m
common
on
one
or
both
dimensions
Y
Cluster
1
X
43 Copyright 2012
43. Prac<cal
Criteria
for
Priori<zing,
Weigh<ng
and
Combining
Dimensions
Ø How
efficiently
they
let
you
divide
the
sample
into
categories
Ø Whether
there
is
a
clear
breakpoint
or
threshold
effect
on
other
variables
Ø Ease
of
defini<on,
measurement
Ø Ease
of
loca<ng
real
representa<ves
when
you
want
to
study
group
in
more
depth
Ø Amount
of
varia<on
on
the
dimension
Ø Amount
of
independent
informa<on
added,
how
much
heterogeneity
the
dimension
accounts
for
Ø Usefulness
as
proxy
for
harder
to
measure
variables
Ø Availability
of
external
informa<on
sources
for
es<ma<ng
prevalence
Ø Power
as
a
predictor
of
differen<al
response
Ø Prac<cal
ability
to
act
differen<ally
depending
on
where
on
the
dimension
people
fall
44 Copyright 2012
45. Demographic
Segmenta<on
Ø Ofen
cri<cized
as
selec<on
criteria
for
usability
studies
Ø But
demographic
variables
have
some
advantages
• Rela<vely
easily
defined,
measured,
detected,
sized
• Easy
to
locate
real
representa<ves
when
• Amount
of
varia<on
can
be
great
• Informa<on
added
at
lile
cost,
makes
them
good
proxies
• Many
products
designed
for
targeted
demographics
• Many
aspects
of
life
may
correlate
with
demographic
dis<nc<ons,
so
can
have
power
as
a
predictor
of
differen<al
response,
needs
• Prac<cal
ability
to
act
differen<ally
toward
them
for
messaging,
sales
channels,
etc.
• First
level
filter
when
you
don’t
yet
know
enough
to
be
more
nuanced
46 Copyright 2012
47. Occupa<onal
Segmenta<on
Ø Profession
Ø Abstract, higher order category (e.g., “knowledge worker,”
“entrepreneur”
Ø General functional area: operations, customer service, finance, IT
Ø Specific roles
Ø Hierarchy: “Executive,” “Manager,” “Supervisor,” “Front line worker”
Ø Context focused: Industry or industry type, company size, business
model, organizational structure
48 Copyright 2012
48. Occupa<onal
Segmenta<on:
Issues
Ø Varying degrees of standardization in nomenclature,
function, and job design
Ø Can your domain knowledge, focus, and sample size
compensate for the “zoom in” problem?
Ø Functional labels can be very difficult to define:
• What is a “knowledge worker”?
• What is a “power user”?
49 Copyright 2012
49. u What
is
a
“knowledge
worker?”
Controller
(finance)
Logis<cs
manager
u
• Focus:
high-‐level
processes
to
• Focus:
tac(cal,
opera(onal
manage
financial
risk
• Priority:
Preven(on
of
low
• Priority:
Increase
efficiency,
probability
events
ensure
smooth
opera(on
• Decisions
based
on
• Needs
quan(ta(ve
data
to
professional
judgment,
manage
processes,
look
for
knowledge
of
best
prac(ce
improvement
opportuni(es
• Sets
policy
for
long
term
• Manages
processes
in
real
(me
50 Copyright 2012
50. Psychographic
Segmenta<on
Ø Abtudes,
preferences,
values
Ø Intended
to
predict
“resonance”
for
messaging
Ø Also
ofen
emphasized
in
personas
for
broad,
generalizable
implica<ons
Ø How
strongly
do
they
relate
to
or
predict
usage
paern
or
other
behaviors?
Ø Are
they
really
more
“stable”
than
behaviors?
Ø How
hard
are
they
to
measure
reliably
and
validly
Ø Self
report
versus
behavioral
self-‐iden<fica<on
51 Copyright 2012
51. Behavioral
Self-‐Iden<fica<on
What
can
you
say
about
psychographics
(e.g.,
preferences)
of
people
who
gather
In
these
venues?
52 Copyright 2012
53. Behavioral
Segmenta<on
Ø Self-‐reported
versus
observed
Ø Purchasing
behaviors
Ø Usage
behaviors:
Amount?
Variety?
Qualita<ve
paern?
Ø Expressed
behavioral
inten<ons:
• How
immediate?
• Evidence
of
preliminary
steps
to
confirm?
Ø Evaluate
degree
of
demonstrated
associa<on
with
behavior
of
ul<mate
interest
54 Copyright 2012
u
54.
Marke<ng
Segments
&
UX
Categories:
The
Ideal
Time
Non-‐
Target
Target
UX
delivers
promised
value
(and
more)
à
sa<sfac<on,
reten<on
Targeted
value
messaging
Purchase
decision
increases
concentra<on
of
process
filters
out
most
poten<al
buyers
of
non-‐target
popula<on
55 Copyright 2012
55.
Marke<ng
Segments
&
UX
Categories:
The
Ideal
Warning:
Time
Ø This
is
most
likely
when
Market
Segmenta<on
and
UX
categoriza<on
map
to
each
other
Ø But
market
segmenta<on
guides
strategies
for
ini<al
filtering,
rather
than
ongoing
experience,
so
relevant
and
available
dis<nc<ons
in
ac<on
may
be
different
Ø UX
has
to
provide
extended
sa<sfac<on
over
a
range
of
encounters
for
each
user
Ø UX
has
more
at
stake
in
each
touch
point,
because
goal
is
engagement
for
an
already-‐filtered
audience
Ø Therefore,
UX
may
introduce
deeper
and/or
transverse
dis<nc<ons
essaging
filters
out
most
of
non-‐target
Targeted
value
m
increases
concentra<on
of
Purchase
decision
process
UX
delivers
promised
value
(and
more)
à
sa<sfac<on,
poten<al
buyers
popula<on
reten<on
56 Copyright 2012
56. Tips
Ø Method
triangula<on:
Ø Start
with
criterion
groups
(differences
you
really
care
about)
and
then
look
for
differen<ators.
Ø Start
with
possible
differen<ators
and
test
to
see
if
they
do
predict
differences
you
really
care
about.
Ø Test
dis<nc<ons
among
segments
that
people
already
believe
in
to
validate
that
they
really
do
predict
something
important
and
ac<onable
Ø Par<al
alignment
on
a
few
variables
of
different
types
may
be
more
robust
and
useful
than
than
op<mizing
for
“clean”
dis<nc<ons
Ø Priori<ze
dimensions
based
on
both
prac<cal
and
conceptual
tradeoffs
57 Copyright 2012
57. More
Tips
Ø Test
dis<nc<ons
across
mul<ple
studies,
or
do
cross-‐
valida<on
within
your
sample
by
splibng
it.
Ø Consider
impact
of
variables
one
at
a
<me
rather
than
only
in
combina<ons,
to
reduce
risk
of
illusory
precision
Ø Try
to
work
within
exis<ng
segments,
but
be
prepared
to
show
how
different
contexts
may
make
transverse
segments
more
or
less
relevant
Ø Studying
pre-‐defined
segments
one
at
a
may
blind
you
to
subgroups
that
are
similar
across
segments-‐include
contras<ng
hypothesized
segments
into
samples
within
or
across
studies
Ø Don’t
expect
the
“average”
differences
of
segments
to
show
up
in
small
samples.
58 Copyright 2012
58. David
A.
Siegel
Ph.D.
Dray
&
Associates,
Inc.
Minneapolis,
MN
USA
david.siegel@dray.com
www.dray.com
+1
612
377
1980