19 Auto Model - Classification Video Lecture | Rapid Miner: Learn and Master - Data & Analytics

[Music]
prototyping is a challenge for beginners
and advanced users alike finding a first
set of good models quickly is what
advanced users want to do and getting to
a suitable model which fits your task is
often the problem for beginners both of
these challenges can be solved with the
auto model feature so that the beginner
can create a model with confidence and
for the advanced user it saves a lot of
time getting started so let's see how it
works the first step is to select some
data you can either choose some data you
have in your repository which is your
local archive or you can import it from
a file or database we will for most of
this demo use one of the samples which
gets shipped with rapid miner studio a
good choice is the Titanic data set
which has 1,300 examples and 12 columns
of different data types with a good mix
of numeric and polynominal attributes
now for the data set as you can guess
from the name we have a collection of
the known information about the
passengers which embarked on the famous
maiden voyage of the RMS Titanic from
England to New York the example set
contains besides personal data like age
gender and number of family members on
board also information about the class
in which they traveled ticket price or
where they boarded the ship now of
course with this data the most
interesting challenge is to try and
model the potential survival of each
passenger and therefore the task at hand
is a classification problem our label or
target variable is obviously the
attribute named survived in this next
step we're asked to identify which value
of the label is for us the important one
in our case we have only two classes and
we're interested in yes you can also
rename the classes or reduce the number
of values by mapping several classes to
the same value
now this next page is particularly
interesting and important every
attribute of our example set is shown
here with a tick box a traffic light and
a quality chart the status colors mean
the following read should most likely be
excluded from the model creation process
Orange make sure to double check if you
really want to select this attribute as
input green this seems to be an
important input for model generation
selecting the right inputs for modeling
is important to make sure you include
every relevant aspect but it's also
important to exclude unimportant
features which only increase the runtime
so make sure to review this
here on the right we have our
information panel which will help to
understand the quality chart better the
quality chart gives you an indication on
correlation between an attribute and the
label the eye deenis the stability and
the fraction of missing values in a
column as mentioned here in general you
should prefer columns with low values
for missing stability and eye dienes a
lot of missing entries will not help to
find patterns for explanation so that's
a no-brainer now you can see why cabin
is flagged to be excluded as clutter
stability is high if the values in an
attribute showing little to no variation
again not helpful for pattern
identification in that case I'd eNOS
tells you how much a column is like a
unique identifier which means each value
appears only once in an attribute and
therefore again is not useful in finding
patterns a random ID like the ticket
number or something as specific as the
name of a passenger clearly is of that
type lifeboat is flagged orange which
tells us that this is something where
rapid minor is unclear if it's a good
idea to select this attribute missing is
high because most people didn't make it
to one and therefore these values are
missing at the same time the correlation
between this attribute and the target
which we want to predict is very high so
we might have a great proxy here our
challenge is to predict if someone is
going to survive assuming the incident
has not happened yet but the information
on lifeboats is a result of the disaster
so due to the causality of the
correlation we have to exclude the
attribute now on this page we get a
pre-selected subset of suitable models
with respect to our task and to the size
of our example set for some of them
regularization or optimization is very
useful to get a better performance so we
can choose to include this automatically
here if you're concerned about the
runtime you can decide to unti it but it
will be best to leave it on besides
creating the models and some helpful
analytics on each we'll also create a
general correlation matrix and an
overview on the importance of the
columns you can turn this off here in
case you're not interested in that but
we won't so we can show you what it
looks like all right now let's hit the
Run button this will take a bit of time
because we're calculating all the models
including validation and optimization by
selecting the overview we can however
look at what has happened so
are already now okay the last entry here
is empty so it is the gradient boosted
tree which is taking some more time
that's not surprising because it is an
ensemble model and therefore is
calculating many different trees in the
table below the graphs we get the run
time of each model and a selection of
performance metrics if you're not sure
which one to look at then we recommend
considering accuracy but also the robust
AUC which stands for area under the
curve as I click on the rock comparison
you can nicely see the curves from which
the AUC is calculated to understand more
about the ROC curves and the different
quality measures please refer to the
tutorials on validation and model
identification respectively now as we
switch back to the overview we see that
so far random forests and is the model
with the highest AUC but also required a
multiple of the other models runtime for
the gain of only about half a percent
higher AUC we had to pay a high price
with the amount of run time the same
applies to our gradient boosted tree
which has just been completed so the
auto model feature provides a great help
here for you to find the most economic
approach between calculation time and
performance before we look into the
details which are provided with each
model let's first look into the general
section up here the data section is
obvious so I'll move on directly to
weights this section helps you to
quickly see which of the attributes were
actually the most relevant for the
models predictions in the case of our
Titanic example clearly the gender is
the most dominant one but it is
immediately followed by monetary
attributes you can see something similar
in the correlations namely that gender
and survival are comparably higher
correlated as well as passenger class
and fare with each other and with
survival okay now here at the bottom we
can see some further helpful details
about each of the models that have been
chosen for the decision tree you get a
nice chart and the full confusion matrix
with all performance details in the
optimal parameters you can see which
parameters if any have been optimized
and what the best values are the last
two entries shown here are lift chart
and simulator the simulator is very
helpful to run through some what-if
scenarios and it includes the option to
ask rapid miner to perform atomization
for example you can ask for the
following optimization how can I
maximize my chance to survive given that
I'm a 40-year old man
who embarks on the voyage in Cherbourg
now you can see that even with optimized
conditions the chance for survival in
this case can only be increased to
slightly above random chance and since
age and gender were predetermined you'd
better make sure to travel in first
class as the biggest supporter of a
positive survival outcome the lift chart
is often used for example during
marketing campaigns and helps to assess
the impact of using a model to identify
campaign targets as compared to just
sending out messages randomly last thing
to demonstrate here is what happens if
you have some examples in your data set
that you want to predict for that we'll
select the customer data which we've
been using also in the other tutorials
check out the video on data loading to
find out how its imported as you can see
the churn label contains some missing
values we will go through the next steps
here rapidly since we've just explained
them and to speed things up even more
will avoid the calculation intensive
gradient boosted tree model while the
remaining models are calculated we can
already see the predictions for our
unlabeled data as made by the naive
Bayes model the green and red colors are
helping you to understand why an example
was classified positive or negative so
this example here was classified as
loyal despite the fact that the customer
is female which otherwise would have
been more an argument for churn but if
you look at the confidence then this is
not a very strong case this example here
in contrast was classified as churn with
all values across the four attributes
supporting this decision with a very
high confidence if you're an experienced
user than one thing you may have been
asking yourself during this demo more
than once is the question what if I want
to do something different than the
options provided for example what if I
want to optimize other parameters in
that case you can simply click down here
on open process and you're free to
change or add anything you like with the
big advantage that you won't need to
start from scratch so Auto model is not
a black box which you can't control but
rather a customizable and transparent
feature to support your productivity
with this we want to end this tutorial
and we thank you very much for watching
[Music]