openproblems-bio / openproblems

Formalizing and benchmarking open problems in single-cell genomics
MIT License
287 stars 76 forks source link

Mean score should be computed & displayed separately for raw or scaled #780

Closed mxposed closed 1 year ago

mxposed commented 1 year ago

On the website currently the Mean score is a mean of raw scores + scaled scores

LuckyMD commented 1 year ago

The mean score should be a mean of the scaled scores only. That's why the scaling was done. I think that is currently the case (at least spot-checking denoising suggests that it is correct).

mxposed commented 1 year ago

I'm sorry if I missed that or if it's already resolved, but I see this is currently the case on the website, and I checked the code for commits and I don't think it's been fixed

LuckyMD commented 1 year ago

Not sure I understand what you mean... spot checking the top performer for DR 5k PBMCs again (https://openproblems.bio/benchmarks/dimensionality_reduction/tenx_5k_pbmc/) confirms that the mean of the scaled values of the 10 metrics is 1.554, which is what is shown. Especially if you go on raw and see the RMSE is 66k, you can see that the mean does not take this value into account.

LuckyMD commented 1 year ago

Or do you mean that there should be a different mean score for raw scores? This is not the case by design as the mean is what is ranked by... and to calculate this mean you need the normalized metric outputs, which is what scaling is.

mxposed commented 1 year ago

Oh, I now see what you mean, sorry, my bad. Yes, it's a bit confusing that by default I see raw scores, but mean score of scaled values. Closing this, my apologies

LuckyMD commented 1 year ago

Good point though about seeing raw vs scaled scores as default. I think we just need a better "tooltip" or explanation of this on the website to clarify.