Sunburst/treemap path

[emmanuelle]

Still WIP, TODO

• write doc
• discuss the different cases we want to support
• discuss case of hover columns

[emmanuelle]

needs work

[nicolaskruchten]

hehe yes :)

[emmanuelle]

oops yes. The PR is not completely ready but I'll still appreciate your thoughts about the API and how it's implemented, then I will polish if this is the right direction, unless there is some refactoring to do first.

[mwouts]

Thank you @emmanuelle for working on this! This will make is way easier for everyone to produce these beautiful plots, and I will certainly be a user of this.

I have a few remarks/questions:

• Would it make sense to could extend this to non-rectangular trees by using 'None' in the path?
• Sometime want to compute aggregates on all nodes, not only for the 'value', but also for the hovertext, like the life expectancy in the plot below. Do you think we could cover this use case with a dataframe? Maybe with intermediate nodes? (I was using a function)
  
• (technical) do you want to add epsilon when you compute the sum, until branchvalues='total' too strict plotly.js#4405 gets fixed? See Sunburst plots only work when the data has type 'int' mwouts/easyplotly#3 for an example that produces an empty plot because of float comparisons issues.

[nicolaskruchten]

Thanks for taking a look @mwouts !

Re intermediate nodes, I think we should try to accept null/None/NaN for trailing path elements somehow, yes.

Re values for intermediate nodes, let's see how far we can get by having users provide them with the approach above. We could accept an aggregating function as a parameter eventually if this is too limiting :)

Also, I don't think we should add an epsilon, as this will be released as part of 4.5.0, which will include the fix for plotly/plotly.js#4405

[emmanuelle]

Thanks for taking a look @mwouts ! @nicolaskruchten you were mentioning that we could accept a dictionary of lambda functions for doing the aggregation, is this what you had in mind?

[emmanuelle]

In order to address the case of lines with None, maybe we should have branchvalues='remainder' as default and when doing the groupby, put the value of the parent to 0 is it's not already specified? (instead of computing the sum as it is implemented now?).

[emmanuelle]

@mwouts we were also looking for a way to address the case of https://plot.ly/python/sunburst-charts/#sunburst-chart-with-a-continuous-colorscale but it seems hard to use a generic API for this...

[emmanuelle]

After all I included an EPS fix because otherwise my doc example (with tips) did not work! On second thoughts, I also kept the groupby mechanism (including computing the sum of children) since if we want to generalize to more complex aggregations we will have a similar syntax, whereas switching to a remainder mode implied a somewhat different code.

[emmanuelle]

I wrote the doc for sunburst but not for treemap since it will be mostly a copy so I'll update after a round of review.

[mwouts]

Thanks! Glad that helps!

@mwouts we were also looking for a way to address the case of https://plot.ly/python/sunburst-charts/#sunburst-chart-with-a-continuous-colorscale but it seems hard to use a generic API for this...

Well don't you think the same approach, generalized to magic underscore notation, could work? In my example (scroll to the second treemap plot) I have defined marker_color as a function, but I believe a column of the dataframe (maybe extended to include the parent nodes) could work as well.

[nicolaskruchten]

Some feedback to start with:

1. The path should go from parent to child rather than the other way around
2. For the continuous-color case, it makes sense to me that for "generated" parent nodes (i.e. ones that don't have a dedicated row whose path ends in None!) for the color of the parent to be the value-weighted average of the children's color. (NB color here is a scalar!)
3. For the discrete-color case, it's trickier: there's no sensible 'implicit' behaviour that I can think of other than imputing some dummy value to color by, or to error out

[nicolaskruchten]

Overall this is really, really awesome though, I love the symmetry with parcoords/parcats/splom !

[nicolaskruchten]

PS: I think we can probably use a similar API for px.sankey

[nicolaskruchten]

As a motivating example for the implicit behaviour in continuous color, think about a case like px.sunburst(gapminder2007, path=["continent", "country"], values="pop", color="lifeExp")

With my scheme, the continent sectors would be colored by the population-weighted average life expectancy, which is pretty much what you'd want. Ditto for values="gdpPercap" etc.

[emmanuelle]

Question about color: if color_continuous_scale is passed but color=None, then a discrete color sequence will be used. Is it ok or should we take the values argument instead? I'd rather keep the current behaviour (not too much magics) but just checking.

[nicolaskruchten]

Yep the current behaviour is fine when color is not passed IMO

[nicolaskruchten]

under what circumstances do we do this, and what's the reasoning?

[emmanuelle]

it's if neither values nor color was passed. I think I need to use some sort of aggregation function because after I call reset_index to get the groups, but maybe it's possible to do it in a more elegant/efficient way.

[emmanuelle]

I could get rid of this part with the count_colname logic.

[nicolaskruchten]

This causes an error: px.sunburst(df, path=["continent", "country"], values="pop", color="continent")

[emmanuelle]

I think I have addressed all your comments!