scale domain/range should not place items in legend if they are not in plot

[jowens]

This was an Altair graph. I wanted to assign colors to vendors using a domain/range Scale and make many graphs with that Scale. Some graphs do not have all vendors. I expected that only the vendors that appeared in the graph would appear in the legend. (Consider an uber-Scale that has hundreds of items in domain/range but only a few appear in any particular graph.)

In the JSON below (VL 2.0.0-beta.11) I see both AMD and NVIDIA in the legend, but only NVIDIA points appear in the graph, so I expect only NVIDIA should be in the legend.

{
  "mark": "point",
  "data": {
    "values": [
      {
        "Launch": "1995-09-01 00:00:00.000000000",
        "SM count": 1,
        "Vendor": "NVIDIA"
      },
      {
        "Launch": "1996-09-01 00:00:00.000000000",
        "SM count": 2,
        "Vendor": "NVIDIA"
      }
    ]
  },
  "encoding": {
    "y": {
      "field": "SM count",
      "type": "quantitative"
    },
    "color": {
      "field": "Vendor",
      "scale": {
        "domain": [
          "AMD",
          "NVIDIA"
        ],
        "range": [
          "#ff0000",
          "#76b900"
        ]
      },
      "type": "nominal"
    },
    "tooltip": {
      "field": "Model",
      "type": "nominal"
    },
    "x": {
      "field": "Launch",
      "type": "temporal"
    }
  }
}

[domoritz]

There is no clean way to do this in Vega as far as I know. I remember that we talked about domain mapping that specifies what colors items in the domain should be mapped to but I can't find the reference right now. Either way, we need support in Vega first so I'm closing this issue here. Please file an issue in Vega first.

[kanitw]

@jowens Thanks for filing an issue. However, conceptually, legend is a visualization of the scale. Thus if you set the scale domain, it should appear on the legend by design. Doing something different is actually violating the concept.

However, a better alternative would to use a formula transform to derive a new field with specific color value and just the data value as the color.
This requires us to allow disabling scale (#2400), but for now you can actually start doing this in Vega by referring to a field without passing it through a scale.

Our previous discussion of a "replace" transform could be handy for this task as well #1761.

[domoritz]

Yes, #1761 was what I was thinking of. Thanks @kanitw!

[jheer]

In Vega, you can control which values appear in a legend using the values property - overriding the scale domain in the process. It's basically a white list. For more, see https://vega.github.io/vega/docs/legends/

[kanitw]

Oh yeah, if you want to manually change values for each plot (or edit the underlying to wire values to a signal based on your dataset) then it works.

What I mention above is more like what could work natively in Vega-Lite in the future :)

[domoritz]

@kanitw Do you think we should wire up things in Vega-Lite to use the original domain of the scale to set the values of the legends? Even if you or I don't implement this right now, we could mark this as "help wanted".

[kanitw]

I dont think so. It will be extremely complicated to make it work correctly with union domain. The requested use case can be achieved cleaner and easier via formula+scale disabling.

…

On Tue, Aug 8, 2017 at 4:28 PM Dominik Moritz ***@***.***> wrote: @kanitw <https://github.com/kanitw> Do you think we should wire up things in Vega-Lite to use the original domain of the scale to set the values of the legends? Even if you or I don't implement this right now, we could mark this as "help wanted". — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#2802 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAGypbqjDDZKLGLOpKPxqLr3exxHBdMiks5sWO8pgaJpZM4OwDoy> .

[kanitw]

Actually I'm a bit wrong -- if we disable scale, then there would be no legend. Hmm, I'll re-open this for now then. Not sure what's the best solution yet.

[jowens]

Thanks for always taking my issues seriously, whether they're good, bad, or weird.

[g3o2]

In the above example, I don't quite understand how it would be more tedious to provide a manual scale not based on the data than a manual values array for the legend. Does this mean that it would be programmatically possible to inject a custom scale but not a custom values array ?

[jowens]

@g3o2 My use case is one defined domain/range scale, many graphs. Consider if I have dozens of vendors and each one gets its own (custom) color, but any individual graph has only a small subset of the total number of vendors.

I can't address your question.

[kanitw]

Closing as it appears to be no clear solution here. I think this is a bit too complicated to support natively in Vega-Lite -- better just do it in Vega or D3.

If you have a good proposal, please feel free to reopen and propose one.