Yesterday's coding

Hi @robertlundberg

Is there a link to yesterday’s live stream, please?

Thanks

Hi @JWalker,
For sure, it’s here: https://www.youtube.com/watch?v=Qy9qxAMFnzc :slight_smile:
(We should maybe start doing a post here for each of the streams so they are easy to find)

Hope you enjoy it!

1 Like

For the first time in ages I made it! Good to see Merge getting a good workout :wink:

And started thinking that instead of a couple of numeric parameters in my redshift estimator I should go back the dB and see if there are telescoping images (small!) I could use instead (just one challenge - the parameters to be superseded by the image are “shape size” - I’d have to dig into how to use CNN’s for ~measurement)

Your comments on the dataset and oddities it contains (39 bathrooms???) just underscore what I heard someone say recently: 70% of a datascience project’s time is spent on wrangling data. Sample of 1 comment, but it has the ring of truth to it!

Haha glad you made it @JulianSMoore! :grin:

Cool idea to use telescoping images, are you thinking in combination with your other data or standalone?
If trained well, I’m fairly sure a CNN could return encodings correlating to the shape size.

That with the data goes back to the age old saying, what goes in comes out (or rather, if shit goes in, shit comes out). A well prepared dataset is the basis for a good model :slight_smile:
I’m still surprised that the model performed decently even with the dataset oddities. We need to add in a SHAP test to make it easier to dig into what had the highest impact on the prediction, would be really interesting in this case.