BUAD
425
1
Trojan
Horse:
Unexpected
Style
at
Your
Door
Case
Questions:
Decision
Tree
1. Using
the
training
data,
build
a
decision
tree
model
to
predict
probability
that
a
member
will
purchase
a
box.
If
you
set
a
cut-‐off
threshold
of
.15,
how
many
members
would
you
target
in
the
training
set?
Which
ones?
There
are
3
leaves
we
target
with
36
+
107
+
189
=
332
members
in
the
training
set.
The
leaves
are:
• R
<
16
&&
ClassicGentleman
>=
1
&&
Rugged
>=2
• R
<
16
&&
ClassicGentleman
>=
1
&&
Rugged
<2
• R
<
16
&&
ClassicGentleman
<
1
&&
R
<
8
We
can
simplify
the
description
of
these
members
to:
m
• R<
16
and
ClassicGentleman
>=
1
er as
• R
<
8
and
ClassicGentleman
<
1
co
eH w
o.
2. What
is
the
R2
of
your
model?
In
simple
words,
what
does
this
mean?
rs e
The
R2
is
10.5%
on
the
training
data,
7.9%
on
the
testing
data.
In
words,
this
ou urc
means
that
the
model
only
explains
about
10%
of
the
variability
in
the
data,
which
doesn’t
seem
like
very
much.
We’ll
see
later,
though,
that
it
still
has
enough
classification
accuracy
to
improve
our
marketing
campaign.
o
aC s
vi y re
3. Examine
each
of
the
split
variables.
Do
they
make
intuitive/business
sense?
Explain
each
one.
• Splitting
on
R
is
intuitive.
Customers
who
haven’t
purchased
in
a
long
time
probably
don’t
like
the
service
and
won’t
purchase
again.
ed d
• Splitting
on
ClassicGentleman
probably
occurs
because
members
who
ar stu
like
“ClassicGentleman”
items
like
the
current
box.
• The
box
is
also
probably
more
on
the
“Rugged”
side,
explaining
why
customers
who
purchase
from
Rugged
are
more
likely
to
buy.
is
• The
final
split
is
somewhat
more
confusing.
Customers
who
haven’t
bought
a
ClassicGentleman
product,
but
have
purchased
something
Th
within
the
last
8
months
do
tend
to
like
the
box.
This
split
merits
further
investigation.
One
possibility,
e.g.,
is
that
members
in
this
box
are
new
to
the
service
and,
hence
haven’t
had
the
chance
to
buy
a
sh
ClassicGentleman
product,
but
would
if
they
were
offered
one.
If
they
had,
they’d
be
part
of
the
other
leaves.
These
splits
also
suggest
that
the
current
box
is
very
“ClassicGentleman”,
but
has
“Rugged”
elements
to
it.
Importantly,
this
description
of
the
box
was
generated
1
This
case
was
developed
for
USC
Marshall’s
BUAD
425
by
Prof.
Arif
Ansari
and
Prof.
Vishal
Gupta.
This study source was downloaded by 100000793680026 from CourseHero.com on 06-16-2021 03:06:17 GMT -05:00
1
https://www.coursehero.com/file/22350914/BUAD-425-Trojan-Hose-SOLUTIONS/
,
BUAD
425
by
the
members
themselves,
not
by
what
the
stylists
think
the
members
would
say.
4. Using
the
testing
data
set
and
a
cut-‐off
threshold
of
.15,
build
the
confusion
matrix
for
this
model.
If
you
were
to
target
at
most
an
additional
48,000
members
using
this
model,
what
do
you
estimate
your
profit
to
be?
See
Excel
spreadsheet.
Predicted
Positive
Negative
Actual
TRUE
66
39
FALSE
258
637
From
the
confusion
matrix,
our
model
picks
about
(66
+
258)
/
(66
+
258
39
+
637)
=
32.4%
of
members.
Of
our
total
member
base,
this
is
32.4%
*
m
500,000
=
101,852
>
48,000.
So
we
will
want
to
target
all
48,000
members.
er as
(Notice,
if
this
number
was
smaller
than
48,000,
we
would
want
to
target
co
fewer
than
48,000
members.)
eH w
o.
The
probability
a
targeted
member
purchases
is
66/(66+258)=
20.4%.
Thus,
rs e
the
expected
profit
per
targeted
member
is:
ou urc
45.5
*
.204
-‐
4
*(1-‐.204)
=
$6.083.
o
Across
our
48,000
targeted
members
this
yields
a
profit
of
$292,000.
aC s
vi y re
5. (More
challenging)
The
threshold
.15
was
somewhat
arbitrary.
Build
a
curve
computing
the
expected
profit
on
the
48,000
members
for
different
values
of
the
threshold.
What
threshold
value
would
you
propose?
ed d
We
use
Excels
Data
Table
function
to
construct
the
table
of
values.
The
curve
of
ar stu
profits
for
different
thresholds
is
below.
is
Th
sh
For
very
small
values
of
the
threshold,
we
target
too
many
customers
that
are
unlikely
to
purchase
the
box
and
lose
money
on
customers
that
don’t.
For
very
This study source was downloaded by 100000793680026 from CourseHero.com on 06-16-2021 03:06:17 GMT -05:00
2
https://www.coursehero.com/file/22350914/BUAD-425-Trojan-Hose-SOLUTIONS/
425
1
Trojan
Horse:
Unexpected
Style
at
Your
Door
Case
Questions:
Decision
Tree
1. Using
the
training
data,
build
a
decision
tree
model
to
predict
probability
that
a
member
will
purchase
a
box.
If
you
set
a
cut-‐off
threshold
of
.15,
how
many
members
would
you
target
in
the
training
set?
Which
ones?
There
are
3
leaves
we
target
with
36
+
107
+
189
=
332
members
in
the
training
set.
The
leaves
are:
• R
<
16
&&
ClassicGentleman
>=
1
&&
Rugged
>=2
• R
<
16
&&
ClassicGentleman
>=
1
&&
Rugged
<2
• R
<
16
&&
ClassicGentleman
<
1
&&
R
<
8
We
can
simplify
the
description
of
these
members
to:
m
• R<
16
and
ClassicGentleman
>=
1
er as
• R
<
8
and
ClassicGentleman
<
1
co
eH w
o.
2. What
is
the
R2
of
your
model?
In
simple
words,
what
does
this
mean?
rs e
The
R2
is
10.5%
on
the
training
data,
7.9%
on
the
testing
data.
In
words,
this
ou urc
means
that
the
model
only
explains
about
10%
of
the
variability
in
the
data,
which
doesn’t
seem
like
very
much.
We’ll
see
later,
though,
that
it
still
has
enough
classification
accuracy
to
improve
our
marketing
campaign.
o
aC s
vi y re
3. Examine
each
of
the
split
variables.
Do
they
make
intuitive/business
sense?
Explain
each
one.
• Splitting
on
R
is
intuitive.
Customers
who
haven’t
purchased
in
a
long
time
probably
don’t
like
the
service
and
won’t
purchase
again.
ed d
• Splitting
on
ClassicGentleman
probably
occurs
because
members
who
ar stu
like
“ClassicGentleman”
items
like
the
current
box.
• The
box
is
also
probably
more
on
the
“Rugged”
side,
explaining
why
customers
who
purchase
from
Rugged
are
more
likely
to
buy.
is
• The
final
split
is
somewhat
more
confusing.
Customers
who
haven’t
bought
a
ClassicGentleman
product,
but
have
purchased
something
Th
within
the
last
8
months
do
tend
to
like
the
box.
This
split
merits
further
investigation.
One
possibility,
e.g.,
is
that
members
in
this
box
are
new
to
the
service
and,
hence
haven’t
had
the
chance
to
buy
a
sh
ClassicGentleman
product,
but
would
if
they
were
offered
one.
If
they
had,
they’d
be
part
of
the
other
leaves.
These
splits
also
suggest
that
the
current
box
is
very
“ClassicGentleman”,
but
has
“Rugged”
elements
to
it.
Importantly,
this
description
of
the
box
was
generated
1
This
case
was
developed
for
USC
Marshall’s
BUAD
425
by
Prof.
Arif
Ansari
and
Prof.
Vishal
Gupta.
This study source was downloaded by 100000793680026 from CourseHero.com on 06-16-2021 03:06:17 GMT -05:00
1
https://www.coursehero.com/file/22350914/BUAD-425-Trojan-Hose-SOLUTIONS/
,
BUAD
425
by
the
members
themselves,
not
by
what
the
stylists
think
the
members
would
say.
4. Using
the
testing
data
set
and
a
cut-‐off
threshold
of
.15,
build
the
confusion
matrix
for
this
model.
If
you
were
to
target
at
most
an
additional
48,000
members
using
this
model,
what
do
you
estimate
your
profit
to
be?
See
Excel
spreadsheet.
Predicted
Positive
Negative
Actual
TRUE
66
39
FALSE
258
637
From
the
confusion
matrix,
our
model
picks
about
(66
+
258)
/
(66
+
258
39
+
637)
=
32.4%
of
members.
Of
our
total
member
base,
this
is
32.4%
*
m
500,000
=
101,852
>
48,000.
So
we
will
want
to
target
all
48,000
members.
er as
(Notice,
if
this
number
was
smaller
than
48,000,
we
would
want
to
target
co
fewer
than
48,000
members.)
eH w
o.
The
probability
a
targeted
member
purchases
is
66/(66+258)=
20.4%.
Thus,
rs e
the
expected
profit
per
targeted
member
is:
ou urc
45.5
*
.204
-‐
4
*(1-‐.204)
=
$6.083.
o
Across
our
48,000
targeted
members
this
yields
a
profit
of
$292,000.
aC s
vi y re
5. (More
challenging)
The
threshold
.15
was
somewhat
arbitrary.
Build
a
curve
computing
the
expected
profit
on
the
48,000
members
for
different
values
of
the
threshold.
What
threshold
value
would
you
propose?
ed d
We
use
Excels
Data
Table
function
to
construct
the
table
of
values.
The
curve
of
ar stu
profits
for
different
thresholds
is
below.
is
Th
sh
For
very
small
values
of
the
threshold,
we
target
too
many
customers
that
are
unlikely
to
purchase
the
box
and
lose
money
on
customers
that
don’t.
For
very
This study source was downloaded by 100000793680026 from CourseHero.com on 06-16-2021 03:06:17 GMT -05:00
2
https://www.coursehero.com/file/22350914/BUAD-425-Trojan-Hose-SOLUTIONS/
