Discussion between Toros91 and KK2491

« first day (14 days earlier) ← previous day next day → last day (74 days later) »

Sam

07:16

Morning @Toros91

Toros91

Hey man

Good Morning

Sam

I've a question for you..

More to gauge your opinion than a question actually.

When dealing with a class imbalance, an option is to under sample the over-represented class as to create a balanced training set

Toros91

yeah what is it?

Sam

This is a common approach. Though it feels a little naive. Wouldn't it be better to sample down the over-represented class to include points closest to the decision boundary of the two classes (assuming a binary classification problem)

Toros91

07:31

yeah that is what ROSE package does

it reduces all the the over-represented class by removing them

Sam

But doesn't it remove at random

Toros91

ROSE removes based on removing the bigger class

*majority

i won't say random completely

for better version of it, we use SMOTE

Sam

Yeh - it's a little different than what I had in mind

Toros91

which does it in a better way

Sam

Which is, remove data points from that class that are at the extremes. We only care about data points which we can gauge a seperation from right?

Therefore, shouldn't there be more of a process involved.

Toros91

07:37

hmmmm so what you have said is right, SMOTE does it like that by removing the frequent occurring cases during undersampling. vice-versa for oversampling

when you carryout such analysis it take a lot of time

in my scenario when I used SMOTE it took a lot of time but the problem is, when it does the generation of values then some of them are negative values

so we need to understand the data after the SMOTE generates the data

Sam

SMOTE seems to have somewhat a better process. But, what happens in the case that the frequent cases are closer to the seperetion boundary. If that was the case, they should definitely be used.

Toros91

hmmmm yes, it will reduce those records by adding some synthetic values to satisy that

Sam

hmm

Toros91

i'm I clearing your doubt?

if not I can explain in a better way

Sam

So, I'm thinking something like this

Imagine the position is based on say the likelihood function.. we don't want to sample down to include points with a high likelihood because these are already well predicted points

They will do nothing for our algorithm to discriminate between points who's likelihood are close together

Toros91

07:51

yeah if you see, based on the 3 green values it will create more synthetic records, it will try to reduce the red records i will start with the outer most

I'm talking with respect to SMOTE

Sam

So SMOTE takes into account the values of those points w.r.t. to other class?

Toros91

with respect to its closeness to the target variable and based on the count we give

« first day (14 days earlier) ← previous day next day → last day (74 days later) »

all rooms