[scikit-learn] sample_weight vs class_weight

Nicolas Hug niourf at gmail.com
Fri Dec 4 05:59:03 EST 2020

Basically passing class weights should be equivalent to passing 
per-class-constant sample weights.

 > why do some estimators allow to pass weights both as a dict in the 
init or as sample weights in fit? what's the logic?

SW is a per-sample property (aligned with X and y) so we avoid passing 
those to init because the data isn't known when initializing the 
estimator. It's only known when calling fit. In general we avoid passing 
data-related info into init so that the same instance can be fitted on 
any data (with different number of samples, different classes, etc.).

We allow to pass class_weight in init because the 'balanced' option is 
data-agnostic. Arguably, allowing a dict with actual class values 
violates the above argument (of not having data-related stuff in init), 
so I guess that's where the logic ends ;)

As to why one would use both, I'm not so sure honestly.


On 12/4/20 10:40 AM, Sole Galli via scikit-learn wrote:
> Actually, I found the answer. Both seem to be optimising the loss 
> function for the various algorithms, below I include some links.
> If, we pass *class_weight* and *sample_weight,* then the final cost / 
> weight is a combination of both.
> I have a follow up question: in which scenario would we use both? why 
> do some estimators allow to pass weights both as a dict in the init or 
> as sample weights in fit? what's the logic? I found it a bit confusing 
> at the beginning.
> Thank you!
> https://stackoverflow.com/questions/30805192/scikit-learn-random-forest-class-weight-and-sample-weight-parameters
> https://stackoverflow.com/questions/30972029/how-does-the-class-weight-parameter-in-scikit-learn-work/30982811#30982811
> Soledad Galli
> https://www.trainindata.com/
> ‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐
> On Thursday, December 3, 2020 11:55 AM, Sole Galli via scikit-learn 
> <scikit-learn at python.org> wrote:
>> Hello team,
>> What is the difference in the implementation of class_weight and 
>> sample_weight in those algorithms that support both? like random 
>> forest or logistic regression?
>> Are both modifying the loss function? in a similar way?
>> Thank you!
>> Sole
> _______________________________________________
> scikit-learn mailing list
> scikit-learn at python.org
> https://mail.python.org/mailman/listinfo/scikit-learn
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.python.org/pipermail/scikit-learn/attachments/20201204/66684009/attachment.html>

More information about the scikit-learn mailing list