[HN Gopher] Explaining ML Models: A Non-Technical Guide to Inter...
___________________________________________________________________
Explaining ML Models: A Non-Technical Guide to Interpreting SHAP
Analyses
Author : sebg
Score : 28 points
Date : 2021-11-05 00:23 UTC (1 days ago)
(HTM) web link (www.aidancooper.co.uk)
(TXT) w3m dump (www.aidancooper.co.uk)
| b9a2cab5 wrote:
| This needs more emphasis on the "correlation is NOT necessarily
| causation" part. In my experience trying to use SHAP and other
| "interpretable ML" techniques for data science analysis, trying
| to identify relevant drivers for a metric based on SHAP-like
| techniques is a recipe for disaster. You'll get results that look
| on the surface to someone that is well versed but not an expert
| in the business like they're legit but to an expert they're
| obvious garbage.
|
| The way I and many others were taught in school is that you get
| some dataset and you throw it into a predictor and get results,
| but what they should've taught is how to find things (among
| hundreds of possible attributes, or even attributes that aren't
| collected yet) you think would be causal predictors first without
| looking at the data and then explore and build a model based on
| that.
___________________________________________________________________
(page generated 2021-11-06 23:02 UTC)