Weka Tutorial 09: Feature Selection with Wrapper (Data Dimensionality) Video Lecture | Weka Tutorial - Data & Analytics

hello welcome to the Waker
machine-learning tutorial in this
tutorial we'll cover a question how can
we select the most effective features
from the set of features we have or this
is called feature vector so in order to
do that we have some feature selection
methods and among very popular methods
the most basic methods are of two kinds
the first one is called wrapper method
and filter method so these are two
methods to select the most effective
features from the set of features you
have for your dataset so in this
tutorial we're going to cover the
wrapper method and in the next tutorial
we'll try to solve the problem with the
with the filter method so in Reaper
method we use a subset evaluator and
that subset evaluator will create all
possible subsets from your feature
vector then it will use a classification
algorithm like naive Bayes or bagging or
j48 graft to induce classifiers from the
features in each subset and it will
consider the subset of features with
which the classification algorithm
performs the best so for example if you
have 10 features the evaluator subset
evaluator will try to find out subsets
every possible subset from those ten
features and for example it produces
three subsets the first one contains
three features the second one contains
five features and the third one contains
seven features so with this first subset
if you select naivebayes
as your graphic classification algorithm
then we thought with the features in
your first subset it will apply the
naive Bayes classification algorithms on
your test instances or training
instances and we'll try to figure out
what the classification accuracy is and
on the basis of the classification
accuracy
of the three subsets it produced from
your feature vector it then presents the
best one the best subset it presents you
the bit subset that produces the best
graphic classification accuracy with the
naive based algorithm and you can see
that I have highlighted all possible
subsets with red because to find a
subset even to find a subset this
evaluator really used a search technique
so the possible subsets that the
evaluator produces from your feature
vector solely depends on the search
technique that it applies so you can use
either random search or any kind of
breadth first search or depth-first
search or a hybrid approach that
integrates both BFS and DFS so today I'm
going to show you how you can do the
represented in order to select features
from your feature vector or pool of
features using the wake of machine
learning workbench so we're going to
click Explorer and for this tutorial I'm
going to take I'm going to consider the
iris data set that's given in the wakeup
installation package so this is our iris
our data set so now I'm going to asila a
I'm going to click the Select attributes
tab here and you can see the first one
the first option here is the attribute
evaluator and the second one is the
search method so this is our subset
evaluator our subset evaluators are
inside inside this option and the
methods to search the subsets we have
the options here you can also choose if
your if you are trying to find the
subsets of features by using the full
training set or by using a
cross-validation and you can also define
the number of folds and see it for the
cross-validation here so now if you take
a look at the data in the pre process
tab you can see that we have a
attribute here that's a nominal nominal
attribute here the type is nominal so we
are trying to classify the instances on
the basis of this number five feature in
our data set
so in select attributes you can see that
it can select so you can select the
class attribute by means you're trying
to separate the instances from each
other so in order to use the wrapper
method we can choose the attribute
evaluator you can see that we have lots
of evaluators and with these evaluators
you can either use wrapper method or you
can use other filter method so for this
usual as I said before this is only
focused for wrapper method so we are
going to choose the third one that's
called classifier subset eval so I
selected that and if you click on this
panel then you can see that we have a
more option and it says that the
classifier subset evaluator evaluates
attribute subsets on training data or a
separate hold of testing set and it uses
a classifier to estimate the merit of a
set of attributes so to estimate the
merit it uses normally it uses the
predictive accuracy or sometimes it also
can use the precision recall and f-score
on the classification task so here you
can define the classifier you are trying
to use on the features of subset of
features that the attribute evaluator is
producing so for this example we are
trying to stick with naive Bayes we
select that and our holdout file is
click to set holdout test instances we
are just keeping it as default so these
are the two options the other one is use
training and yes we are saying that okay
use the training data to get the best
subset from the subset produced by
activity
we are clicking in okay and in search
method so you can see that we add the
default is selected as best first the
best versus hybrid approach of both BFS
and DFS and you can see that we have
some exhaustive searches genetic search
really stepwise linear for selection
ranker and so on so you have numbers of
options to select as a search method so
that depends on your classification task
or that also depends if you have time
and memory then you can use a BFS
approach BFS search technique and if you
want to utilize your time and effort
then you can use a DFS search technique
so it all depends on you and your task
so now I'm just clicking on start and
you can see that the selected attributes
are 1 3 & 4 so if you go to the pre
process tab you can see that the two
that the simple width according
according to 9 base it says that with
these three features we can we can
classify the instances in the data set
is data set best with the naive Bayes
classification algorithm also you can if
you can if you can also set the
classifier as a different one for
example instead of naive Bayes we're
trying to find it with a bagging the
classification algorithm and again if
you click on start then you can see that
bagging is saying that the most
effective features are number 3 & 4 so 9
pace is giving you 1 3 4 and bagging is
giving you 3 & 4 this means that the
performance on classification
performance of naive Bayes on this
particular data set called iris is best
when you were selecting 1 3 4 and you're
omitting the second feature in in your
feature set or feature pool or feature
vector whatever you say so we can do it
right now so a code
to this information we use night bass
and we have seen that the best features
are 1 3 & 4 so we can select 1 3 4 & 5
we just invert it so we are just
dominating number 2 from our feature set
and we can go to the classify tab you
can choose the classifier to night base
and we get the results here so you can
see the hundred percent position 100%
recall for one class the second class
say is ninety two point ninety two
percent position 96 percent recall and
the third class is a 95 percent position
and only two percent recall so this is
guaranteed that with those three
features from our four features
excluding the class feature gives you
the best result if you just select it
you undo that and select any other for
example if you want to try with two
three and four and you're just omitting
number one feature then you exclude that
go to classify tab and start again so
you can see that we are currently this
one was done by the selected features
the most effective features and this was
done by randomly selecting a feature to
be omitted from your fixture feature
vector so you can see that the precision
recall and F score has dropped so this
one says that
okay I'm guaranteeing you dad with that
with those three features by applying
ninety base you get the best accuracy
result so I think that's it for today
and bye-bye this you can always use the
select attribute stabbing wake Explorer
to see to find out the best possible
subset of features if you're using
subset evaluator and using a wrapper
method for subset of feature selection
thank you