vega
Support Contour Plot
It is related but fundamentally different than a binned chart - kernel density estimation (KDE) reconstructs a continuous distribution from observed discrete samples. Vega 3 supports it, as well as underlying mechanisms for fitting and plotting other continuous theoretical distributions (e.g., normal). These could be useful for EDA. See the violin plots example in the Vega 3 repo for one use case that Altair users (for example) might benefit from.
I experimented with contour plots a bit last summer. While they look great, there are some challenges with overlapping contours if the data is faceted. I could not even use pure Vega when I wanted multiple contours to share the same color scale. The issue is that we don't compute the density independently from the lines.
Not a bug, but perhaps a feature request, along with detailed descriptions of the required functionality. IIRC this might require an additional operation that calculates levels for each faceted contour and then performs aggregation to determine a scale range?
Yes, this is not a bug in Vega but a missing feature. Contour plots in Vega use a single transform that computes density and the contour lines in one step. If we break these steps apart, we can compute the range of densities and share the same color scale across contours. There may be more to this (as well as some design work on the Vega-Lite side) so that I don't feel confident to file an issue with Vega yet.
The density + contour calculation is "outsourced" to d3-contour, which computes the density and then outputs contours as GeoJSON: https://github.com/d3/d3-contour/blob/master/src/density.js
This method internally performs KDE and invokes contour generation. As is helpful, one might contribute a PR that breaks those steps apart into subroutines, e.g., one that produces a blurred (KDE) value grid and one that produces shared contour threshold values given one or more grids.
Ideally Vega would avoid internal duplication of effort, e.g., a collection of contours might be created once via a contour-specific groupby aggregate transform and then faceted by groupby keys.
Would the faceting issue be a problem for violin/density plot as well?
Note that I'm adding a separate issue for violin / density plot in https://github.com/vega/vega-lite/issues/3442, and keep this issue mainly for contour plot.
@zening just mentioned today that we can "hack" and use contour to highlight point as a way to annotate plots too. (She will post a screenshot.)
For my consistency project, I sometimes want to highlight a cluster of marks. For example, in the mockup below, my "console" prints a warning that says the blue in the color palette of view 2 conflicts with the blue in view 1. When the user hovers on the text warning, I want to highlight all marks in view 1, and the blue in view 2 palette. (Ideally I also want to draw a line connecting the two areas in conflict to make it explicit that the warning concerns two areas.)
To highlight the mark cluster in view 1, I want to draw a "cloud" that loosely fit the shape of the mark cluster. I tried to hack the vega contour example to do this:
{
"$schema": "https://vega.github.io/schema/vega/v3.json",
"width": 500,
"height": 400,
"padding": 5,
"autosize": "pad",
"data": [
{
"name": "source",
"url": "data/cars.json",
"transform": [
{
"type": "filter",
"expr": "datum['Horsepower'] != null && datum['Miles_per_Gallon'] != null"
}
]
},
{
"name": "contours",
"source": "source",
"transform": [
{
"type": "contour",
"x": {"expr": "scale('x', datum.Horsepower)"},
"y": {"expr": "scale('y', datum.Miles_per_Gallon)"},
"size": [{"signal": "width"}, {"signal": "height"}],
"thresholds": [0, 0.001, 0.002, 0.003, 0.004, 0.005, 0.006, 0.007, 0.008]
}
]
}
],
"scales": [
{
"name": "x",
"type": "linear",
"round": true,
"nice": true,
"zero": false,
"domain": {"data": "source", "field": "Horsepower"},
"range": "width"
},
{
"name": "y",
"type": "linear",
"round": true,
"nice": true,
"zero": false,
"domain": {"data": "source", "field": "Miles_per_Gallon"},
"range": "height"
},
{
"name": "color",
"type": "sequential",
"zero": true,
"domain": {"data": "contours", "field": "value"},
"range": ["white", "#3ae0ae", "white", "white", "white", "white", "white", "white", "white"]
}
],
"marks": [
{
"type": "path",
"from": {"data": "contours"},
"encode": {
"enter": {
"stroke": {"value": "#888"},
"strokeWidth": {"value": 0},
"fill": {"scale": "color", "field": "value"},
"fillOpacity": {"value": 0.35}
}
},
"transform": [
{ "type": "geopath", "field": "datum" }
]
},
{
"name": "marks",
"type": "symbol",
"from": {"data": "source"},
"encode": {
"update": {
"x": {"scale": "x", "field": "Horsepower"},
"y": {"scale": "y", "field": "Miles_per_Gallon"},
"size": {"value": 4},
"fill": [
{"test": "true", "value": "gray"},
{"value": "transparent"}
]
}
}
}
]
}
Can hack contour to highlight a single data point (or a subset of points) as well:
{
"$schema": "https://vega.github.io/schema/vega/v3.json",
"width": 500,
"height": 400,
"padding": 5,
"autosize": "pad",
"data": [
{
"name": "source",
"url": "data/cars.json",
"transform": [
{
"type": "filter",
"expr": "datum['Horsepower'] != null && datum['Miles_per_Gallon'] != null"
}
]
},
{
"name": "contours",
"source": "source",
"transform": [
{
"type": "contour",
"x": {"expr": "scale('x', 132)"},
"y": {"expr": "scale('y', 32.7)"},
"size": [{"signal": "width"}, {"signal": "height"}],
"bandwidth": 30,
"thresholds": [0.1, 0.2, 0.3, 0.4, 0.5]
}
]
}
],
"scales": [
{
"name": "x",
"type": "linear",
"round": true,
"nice": true,
"zero": false,
"domain": {"data": "source", "field": "Horsepower"},
"range": "width"
},
{
"name": "y",
"type": "linear",
"round": true,
"nice": true,
"zero": false,
"domain": {"data": "source", "field": "Miles_per_Gallon"},
"range": "height"
},
{
"name": "color",
"type": "sequential",
"zero": true,
"domain": {"data": "contours", "field": "value"},
"range": ["#3ae0ae", "white", "white"]
}
],
"marks": [
{
"type": "path",
"from": {"data": "contours"},
"encode": {
"enter": {
"stroke": {"value": "#888"},
"strokeWidth": {"value": 0},
"fill": {"scale": "color", "field": "value"},
"fillOpacity": {"value": 0.35}
}
},
"transform": [
{ "type": "geopath", "field": "datum" }
]
},
{
"name": "marks",
"type": "symbol",
"from": {"data": "source"},
"encode": {
"update": {
"x": {"scale": "x", "field": "Horsepower"},
"y": {"scale": "y", "field": "Miles_per_Gallon"},
"size": {"value": 4},
"fill": [
{"test": "true", "value": "gray"},
{"value": "transparent"}
]
}
}
}
]
}
Maybe a temporary solution could be implemented? In a separate library, like vega-lite-contour?
So people would be able to start working with it and it would be easier to find out use cases and how to better integrate it into vega-lite in the future?
Pseudo-code
import { add_contour_plot_curves } from 'vega-lite-contour'
import { vega_lite } from 'vega-lite'
const table = [...] // data table
const table_with_contour = add_contour_plot_curves(table, { ... density estimation options ... })
vega_lite.plot(table, plot_options)
