Make frequencies tend to 0 in absence of data

[rneher]

No description provided.

[rneher]

@jameshadfield, can one redeploy this? I increased the padding value for the frequencies a bit which should improve things. But would be good to flip through various builds and play with it a bit.

[trvrb]

@rneher: You should be able to do this through Heroku. Go here https://dashboard.heroku.com/teams/nextstrain/apps, click on "Auspice" and then find the branch you want to deploy and click "Create Review App". I've just done this for explore-normalization-fix.

[rneher]

Thanks, i keep forgetting how this works for auspice. With the increased padding parameter, the transition to 0 happens a little earlier and a little less abrupt:

[rneher]

I do think this is an improvement compared to disallowing normalization. But an explicit grey out would still be better.

[rneher]

By changing minVal

https://github.com/nextstrain/auspice/pull/1325/files#diff-522f919569fe474196547277ab2bf9a235378aa2b41e86b7982aeb79b323cc70R93

one can move the place where frequencies drop to 0 around and also sharpen the transition (the smaller the value, the earlier and sharper the transition is). This is essentially due to competition between the Gaussian smoothing and the constant minVal. For smaller minVal, the Gaussian dominates for longer.

[trvrb]

Thanks Richard. Just clicking around here I noticed: https://auspice-explore-normali-hknihq.herokuapp.com/ncov/global?f_country=Switzerland

I can't quite tell what's going on with the June 2020 timepoints. And also the slow rise of total frequency in Spring 2020 where there are 7 pivots that don't total 100%.

But maybe the path forward is to do something like this and then place a gray box over pivots that don't have >99% total frequency when normalized? But I would generally prefer a sharper transition. Here we have situations like https://auspice-explore-normali-hknihq.herokuapp.com/ncov/global?c=country&f_country=Vietnam where the whole second peak is less than 97% total frequency. I definitely expected "normalized" frequencies to either sum to 100% or sum to nothing.

[rneher]

yeah, this up and down is not good. I changed it to either fully normalized or zero.

[trvrb]

Awesome! Much improved. Thanks Richard. I think I'm okay merging this and then doing a separate PR for "gray boxes". Although it looks like there's a test failing due to linting error.

[rneher]

I fixed the linting issue... @jameshadfield, any other comments?

[trvrb]

I think this is a definite improvement and would be in support of merging.