Is it possible to use Snorkel for Regression problem?

[ff-priya]

In my problem output is an integer number. I have some rules by which I can say about the possible range in which that number will be. So for some input if actual output is 1000. I have some rules by which I can say it should be between 800 - 1200.

I have lot of data where I can say about approximate value of output but have a very small data-set that gives me exact value of output. I can make some rules that will give range in which output value will lie.
Can I use Snorkel to do so?

[ajratner]

Hi @ff-priya,

If you are asking about whether Snorkel can do a traditional regression problem, as your title suggests (i.e. predict values in some continuous range), the answer is no, but this is a direction of future work we are pursuing, so stay tuned!

If you are asking whether you could have Snorkel output discretized ranges of values (e.g. {"800-1200", "1200-1600"}), then yes, seems like you could indeed do that!

Hope this helps,
Alex

[github-actions]

This issue is stale because it has been open 30 days with no activity. Remove stale label or comment or this will be closed in 5 days.

[rjurney]

@ajratner do you have a sense of how snorkel might approach this problem? Both in the sense of writing LFs for buckets atm as well as in snorkel supporting regression in the future? I guess LFs would return continuous values if Snorkel supported this? At the moment it seems challenging to map LFs to ranges... I guess exploratory data analysis will turn up something?

[panagiotis-c]

Supposing one wants to have a float between 0 and 1 as output, I am wandering if this would make sense: assign labels 0 and 1 with LFs like binary classification, and after training the model, using the attached confidence scores (with return_probs) as the continuous variable output.
This way you can't have LFs return float, so you lose some accuracy you could provide there, but maybe with enough LFs this makes sense.

[annemariet]

If you are asking about whether Snorkel can do a traditional regression problem, as your title suggests (i.e. predict values in some continuous range), the answer is no, but this is a direction of future work we are pursuing, so stay tuned!

Hello! Are there any news about this topic?