SAS Visual Statistics Logistic Regression Demo

Logistic regression analysis can help you better understand conversion behaviors of website visitors.

REGISTER FOR A FREE TRIAL OF SAS VISUAL STATISTICS
SAS® Visual Statistics helps you explore data and quickly build and refine predictive models so you can zero in on the right answers to your biggest business questions.

Once you register, you can:
•Gain unlimited access to the technology.
•Start building and refining models on the fly.
•Find more value in your data.

http://www.sas.com/en_us/offers/14q3/visual-stats-try-before-you-buy.html


SUBSCRIBE TO THE SAS SOFTWARE YOUTUBE CHANNEL
http://www.youtube.com/subscription_center?add_user=sassoftware

ABOUT SAS
SAS is the leader in business analytics software and services, and the largest independent vendor in the business intelligence market. Through innovative solutions, SAS helps customers at more than 75,000 sites improve performance and deliver value by making better decisions faster. Since 1976 SAS has been giving customers around the world The Power to Know.®

VISIT SAS
http://www.sas.com

CONNECT WITH SAS
SAS ► http://www.sas.com
SAS Customer Support ► http://support.sas.com
SAS Communities ► http://communities.sas.com
Facebook ► https://www.facebook.com/SASsoftware
Twitter ► https://www.twitter.com/SASsoftware
LinkedIn ► http://www.linkedin.com/company/sas
Google+ ► https://plus.google.com/+sassoftware
Blogs ► http://blogs.sas.com
RSS ►http://www.sas.com/rss

---

Closed Caption:


Hi.
My name is Suneel Grover,
principle solutions architect
for SAS, and today I'll
be sharing a demonstration
using Visual Statistics to
showcase logistic regression
analysis for a variety
of predictive marketing
applications.
With that said,
let's jump into it.
In this demonstration, we'll be
highlighting Visual Statistics
in the context of logistic
regression analysis
for the purposes of
understanding drivers
of website visitors and what
causes them to convert or not
convert.
Over here in the
right menu, you'll
notice that I've assigned
an attribute that represents
a conversion indicator--
if a visitor did or did not
meet that criteria on their
visitation experience.
I now want to add
additional attributes
in the form of classification
and continuous effects
to help understand and
isolate those signals
unique to conversion behavior.
First and foremost as I come
over here to my left menu
where all my attributes
exist, I want
to search for attributes that
represent visitor interest when
they came to the site.
What types of contextual
areas of the site experience
were they checking out?
I'm going to select
those attributes
and fold them into the
analysis and they'll
be assigned as
classification effects.
The next attribute,
let's grab something
that's popular in the web
analytic world in the form
of visitor engagement.
And let's drop that
into the analysis.

Next I want to highlight
a very powerful feature
in Visual Statistics.
What I'll be pointing you to
is how to assign attributes
to this group by role.
And what that means is rather
than building one regression
model on the entire
audience, I now
want to build multiple
regression models
based on the attribute that I
fulfill in this group by role.
I'm going to select a variable
by the name of traffic source
which represents which
marketing channel did
that visitor come from.
As I select it and I assign
it to the group by role,
Visual Statistics
is very efficiently
going to create seven
unique regression models
for each particular channel.
And we can see that over
here in our fit summary plot.
Let's expand that.
So what's happening here?
We've built a model
for social media,
visitors, for those that
originated from organic search,
from blogs, from online
display, and et cetera.
And if I want to explore
what variables are uniquely
important to each of
those particular channels,
I can come over here, I can
select a variable-- let's
select something like those
who are interested in business
analytics-- and I can see by
the horizontal green lines
that those who
originate from blogs,
those who came from paid search,
and those who came directly
to our site, this particular
contextual interest area
is important to them.
But if we select a
different attribute-- let's
say we're interested visitor
interest in Visual Analytics--
the variable importance
plot will update and provide
clear indication as to
which unique traffic
channels showcase interest
in that particular content.
Let's go ahead and
minimize this plot
and let's take a look
at the assessment plot.
For an analyst that wants to
understand if he or she is
building a good model,
Visual Statistics
provides three visualizations
in the context of assessment.
First a lift chart,
second ROC plot,
and third a
misclassification rate plot.

With respect to
outlier analysis,
we can actually
investigate the residuals.
And if we find certain
observations are deviating
outside of our level
of satisfaction,
we can actually select
those offensive observations
and make the decision if we
would like to exclude them
for the analysis or not.

With that said, there are other
property settings to highlight.
If you would like to deal
with impunity missing values,
if you would like to
use variable selection,
and other common techniques
to further optimize
your regression
model, you'll see
a wealth of options over here
in the property settings.

Now how would we take
action on this model
if we had come to the
conclusion that we
were happy with the analysis?
There are a number of
options to take advantage of.
First and foremost, we
could export the score code.
By clicking here,
Visual Statistics
will auto generate
that score code
and we could now bake this
code into a marketing workflow
so that we could operationalize
or productionalize
on this information
and this insight.

Another option we could consider
is deriving predicted values.
And this is a really wonderful
feature in logistic regression.
Once we're done
with the analysis,
we can actually enrich our
existing data set by scoring
every single visitor,
every single observation
in the data set with the
predicted value this regression
analysis produced.
Simply by clicking
on this button,
Visual Statistics will ask us
to either accept the default
naming convention or provide
unique names for these two
attributes.
I'm going to go
with the defaults.
And once I click
OK, you'll notice
that here-- and towards
the bottom of my list-- two
new attributes have been
added to the data set.

Now there's one last
feature I'd like
to highlight before
ending this demonstration
and it comes in the context
of model comparison.
So let's exemplify that we
built this logistic regression.
And now we want to compare
it with a decision tree
analysis or other
modeling analysis
we've done in Visual Statistics
and we want understand which
modeling algorithm does
the best job of fitting
the data and the marketing
question that we're addressing.
So I'm going to come
to the top of my menu
and I'm going to select
model comparison.

Visual Statistics will
open a window that
will allow me to select
each model that I've built.
And I'm going to select
a tree and the regression
for this example.
I'm going to ask Visual
Statistics to compare the two.
What I want Visual
Statistics to inform me of
is which model is better
at fitting the same data.
In this case, it
very quickly tells me
that based on the two
models that I've built
that the tree is doing better.
Now for those of you that are
comfortable in data mining,
you'll notice there are a number
of different fit statistics
that we can do
this comparison on.
We can also adjust
our prediction cut off
and by doing so, we may
actually get different results.

Again we also have access
to assessment plots
that will showcase not just one
model but both models in that
assessment view so that we can
compare and contrast visually.
Now that we understand
which modeling technique is
a more accurate fit
of the data, I'd
like to conclude by summarizing
what we've seen in this demo.
We started by fitting a logistic
regression to clickstream data
and identifying drivers
of conversion behavior
from visitors.
We then showcased capabilities
in further optimizing the model
by using group by functionality,
which allowed us to build seven
different regression models
based on the visitor traffic
channel that he or
she attributed from.
Lastly we showcased capabilities
in comparing regression
modeling with other
modeling techniques
to understand which
actual algorithm is
most accurate in fitting the
data that we're researching.
So that was a demonstration
of Visual Statistics
for predictive marketing
applications using
logistic regression analysis.
We really appreciate
your time today.
We hope you enjoyed it.