Add Legends To Altair Plots

[O957]

For the full scope of this PR, please refer to issue #117 .

Specifically, this PR adds:

• Legends to the plot with labels for forecasts and observations, when both are present.

[O957]

This is not gated on #125 and #18 but it would be nice if these were complete before proceeding further on this PR.

[O957]

On faceted charts, do we want the legend under each plot? By default, if not under each plot, the legend appears all the way at the bottom of the faceted plots.

Details

[dylanhmorris]

I prefer, in order:

1. One legend for all subplots located above the top-most subplot
2. One legend for each subplot located above that subplot
3. One legend for each subplot located below that subplot (as depicted in @AFg6K7h4fhy2's post)

[O957]

The problem I am having is when I do independent domain and range for alt.Color encoding arguments, the outcome is that, when both files are uploaded, the Observations color is overriden and turns from limegreen to steelblue. When I use the system I currently have in place, all labels are their proper color (well, the CI labels to not seem to be the correct opacity) but appear even if the data is not present (Observations legend label present even when just uploading the forecast file).

Details

I tried querying AI somewhat on this but it wasn't able to help really and propered solutions that did not work / were seemingly not sensible.

I will keep at it.

[damonbayer]

The problem I am having is when I do independent domain and range for alt.Color encoding arguments, the outcome is that, when both files are uploaded, the Observations color is overriden and turns from limegreen to steelblue. When I use the system I currently have in place, all labels are their proper color (well, the CI labels to not seem to be the correct opacity) but appear even if the data is not present (Observations legend label present even when just uploading the forecast file).

I tried querying AI somewhat on this but it wasn't able to help really and propered solutions that did not work / were seemingly not sensible.

I will keep at it.

The legend doesn't really add much to the plot, in my opinion. I think we should abandon the PR, if it is proving difficult.

[damonbayer]

I don't think there is any way for the transparencies to match the legend. By definition the 50% CI is overlayed on top of the the 90%, etc. The legend will have no concept of this.

We could swap to different colors for each band to resolve this. To mimic the R plots, use #DEEBF7, #9ECAE1, and #3182BD.

[O957]

We could swap to different colors for each band to resolve this. To mimic the R plots, use #DEEBF7, #9ECAE1, and #3182BD

Sounds great.

This commit 7185da2 was clean and had Forecast in steelblue wo/ actual CI labels, but what you just suggested sounds better.

[damonbayer]

@AFg6K7h4fhy2 I suggest getting the colors programmatically with

>>> import colorbrewer
... colorbrewer.Blues[3]
[(222, 235, 247), (158, 202, 225), (49, 130, 189)]

[damonbayer]

MAX_NUM_CIS is not used to limit ci_pairs. Seems like an oversight.

[damonbayer]

This approch, using rgb_strings, feels more straightforward than division + hex conversion.

import altair as alt
import polars as pl
import colorbrewer

rgb_tuples = colorbrewer.Blues[3]
rgb_strings = [f"rgb({r},{g},{b})" for r, g, b in rgb_tuples]

df = pl.DataFrame({
    'x': [1, 2, 3],
    'y': [4, 5, 6]
})


chart = alt.Chart(df).mark_point(
    color=rgb_strings[2],
    size=100
).encode(
    x='x',
    y='y'
)

chart

[damonbayer]

Again, let's use format_percent from babel.

[damonbayer]

Structure overall feels a bit off.

What about

1. We haveget_available_ci(list_of_quantiles) which returns all of the available ci's (width only, or width, lower, and upper)
2. Then somewhere later we have the availability to subsample the available ci's (either by giving the list of widths we want or by limiting to some max number).
3. Then we get the colors based on the length of the subsampled list of ci's.
4. Then we make the plot with legend.

[O957]

I will try this out.

[damonbayer]

"Get the colors" step may be obviated by https://vega.github.io/vega/docs/schemes/#blues

[damonbayer]

Rather than colorbrewer, we should try to take advantage of https://vega.github.io/vega/docs/schemes/#blues

[damonbayer]

The output looks good, but I think we can make the internals better.

[dylanhmorris]

I thought the plan was to use vega's built-in implementation of the blues palette. Is there some reason that doesn't work? https://vega.github.io/vega/docs/schemes/#blues

[O957]

Visually, this is what the legend looks like if observations alone are uploaded:

Details

and if the forecast file is uploaded after the observations file:

Details

One noted issue is that the color for Observations legend-label changes and is affected by the colors of the Forecasts legend-labels.

[O957]

When you specify scheme="blues", altair applies the blues color scheme to ALL items in the domain, which included, necessarily, the "Observations" label. To get the colors mixed it seems you need to specify range=, which does not have a built in "blues" or color scheme. range and scheme are mutually exclusive (see https://vega.github.io/vega-lite/docs/scale.html). For programmatic access of altair blues (rather than me using the altair blues defined in ), see this relevant but unresolved issue: vega/altair#2617.