View Source Plotting in EXGBoost

Run in Livebook 
Mix.install([
  {:exgboost, "~> 0.5"},
  {:scidata, "~> 0.1"},
  {:kino_vega_lite, "~> 0.1"}
])

# This assumed you launch this livebook from its location in the exgboost/notebooks folder

Introduction

Much of the utility from decision trees come from their intuitiveness and ability to inform dcisions outside of the confines of a black-box model. A decision tree can be easily translated to a series of actions that can be taken on behalf of the stakeholder to achieve the desired outcome. This makes them especially useful in business decisions, where people might still want to have the final say but be as informed as possible. Additionally, tabular data is still quite popular in the business domain, which conforms to the required input for decision trees.

Decision trees can be used for both regression and classification tasks, but classification tends to be what is most associated with decision trees.

This notebook will go over some of the details of the EXGBoost.Plotting module, including using preconfiged styles, custom styling, as well as customizing the entire vidualization.

Plotting APIs

There are 2 main APIs exposed to control plotting in EXGBoost:

Regardless of which API you choose to use, it is helpful to understand how the plotting module works (althought the higher-level API you choose to work with the less important it becomes).

Implementation Details

The plotting functionality provided in EXGBoost is powered by the Vega JavaScript library and the Elixir VegaLite library which provides the piping to interop with the JavaScript libraries. We do not actually much use the Elixir API provided by the Elixir VegaLite library. It is mainly used for the purposes of rendering.

Vega is a plotting library built on top of the very powerful D3 JavaScript library. Vega visualizations are defined according to the respective JSON Schema specification. Vega-Lite offers a reduced schema compared to the full Vega spec. EXGBoost.Plotting leverages several transforms which are not available in the reduced Vega-Lite schema, which is the reason for targeting the lower-level API.

For these reasons, unfortunately we could not just implement plotting for EXGBoost as a composable Vega-Lite pipeline. This makes working synamically with the spec a bit more unwieldly, but much care was taken to still make the high-level plotting API extensible, and if needed you can go straight to defining your own JSON spec.

Setup Data

We will still be using the Iris dataset for this notebook, but if you want more details about the process of training and evaluating a model please check out the Iris Classification with Gradient Boosting notebook.

So let's proceed by setting up the Iris dataset.

{x, y} = Scidata.Iris.download()
data = Enum.zip(x, y) |> Enum.shuffle()
{train, test} = Enum.split(data, ceil(length(data) * 0.8))
{x_train, y_train} = Enum.unzip(train)
{x_test, y_test} = Enum.unzip(test)

x_train = Nx.tensor(x_train)
y_train = Nx.tensor(y_train)

x_test = Nx.tensor(x_test)
y_test = Nx.tensor(y_test)

Train Your Booster

Now go ahead and train your booster. We will use early_stopping_rounds: 1 because we're not interested in the accuracy of the booster for this demonstration (Note that we need to set evals to use early stopping).

You will notice that EXGBoost also provides an implementation for Kino.Render so that EXGBoost.Boosters are rendered as a plot by default.

booster =
  EXGBoost.train(
    x_train,
    y_train,
    num_class: 3,
    objective: :multi_softprob,
    num_boost_rounds: 10,
    evals: [{x_train, y_train, "training"}],
    verbose_eval: false,
    early_stopping_rounds: 1
  )

You'll notice that the plot doesn't display any labels to the features in the splits, and instead only shows features labelled as "f2" etc. If you provide feature labels during training, your plot will show the splits using the feature labels.

booster =
  EXGBoost.train(x_train, y_train,
    num_class: 3,
    objective: :multi_softprob,
    num_boost_rounds: 10,
    evals: [{x_train, y_train, "training"}],
    verbose_eval: false,
    feature_name: ["sepal length", "sepal width", "petal length", "petal width"],
    early_stopping_rounds: 1
  )

Top-Level API

EXGBoost.plot_tree/2 is the quickest way to customize the output of the plot.

This API uses Vega Marks to describe the plot. Each of the following Mark options accepts any of the valid keys from their respective Mark type as described in the Vega documentation.

Please note that these are passed as a Keyword, and as such the keys must be atoms rather than strings as the Vega docs show. Valid options for this API are camel_cased atoms as opposed to the pascalCased strings the Vega docs describe, so if you wish to pass "fontSize" as the Vega docs show, you would instead pass it as font_size: in this API.

The plot is composed of the following parts:

  • Top-level keys: Options controlling parts of the plot outside of direct control of a Mark, such as :padding, :autosize, etc. Accepts any Vega top-level top-level key in addition to several specific to this API (scuh as :style and :depth).
  • :leaves: Mark specifying the leaf nodes of the tree
  • :splits Mark specifying the split (or inner / decision) nodes of the tree
  • :yes
  • :no

EXGBoost.plot_tree/2 defaults to outputting a VegaLite struct. If you pass the :path option it will save to a file instead.

If you want to add any marks to the underlying plot you will have to use the lower-level EXGBoost.Plotting API, as the top-level API is only capable of customizing these marks.

Top-Level Keys

EXGBoost supports changing the direction of the plots through the :rankdir option. Avaiable directions are [:tb, :bt, :lr, :rl], with top-to-bottom (:tb) being the default.

EXGBoost.plot_tree(booster, rankdir: :bt)

By default, plotting only shows one (the first) tree, but seeing as a Booster is really an ensemble of trees you can choose which tree to plot through the :index option, or set to nil to have a dropdown box to select the tree.

EXGBoost.plot_tree(booster, rankdir: :lr, index: nil)

You'll also notice that the plot is interactive, with support for scrolling, zooming, and collapsing sections of the tree. If you click on a split node you will toggle the visibility of its descendents, and the rest of the tree will fill the canvas.

You can also use the :depth option to programatically set the max depth to display in the tree:

EXGBoost.plot_tree(booster, rankdir: :lr, index: 4, depth: 3)

One way to affect the canvas size is by controlling the padding.

You can add padding to all side by specifying an integer for the :padding option

EXGBoost.plot_tree(booster, rankdir: :rl, index: 4, depth: 3, padding: 50)

Or specify padding for each side:

EXGBoost.plot_tree(booster,
  rankdir: :lr,
  index: 4,
  depth: 3,
  padding: [top: 5, bottom: 25, left: 50, right: 10]
)

You can also specify the canvas size using the :width and :height options:

EXGBoost.plot_tree(booster,
  rankdir: :lr,
  index: 4,
  depth: 3,
  width: 500,
  height: 500
)

But do note that changing the padding of a canvas does change the size, even if you specify the size using :height and :width

EXGBoost.plot_tree(booster,
  rankdir: :lr,
  index: 4,
  depth: 3,
  width: 500,
  height: 500,
  padding: 10
)

You can change the dimensions of all nodes through the :node_height and :node_width options:

EXGBoost.plot_tree(booster, rankdir: :lr, index: 4, depth: 3, node_width: 60, node_height: 60)

Or change the space between nodes using the :space_between option.

Note that the size of the accompanying nodes and/or text will change to accomodate the new :space_between option while trying to maintain the canvas size.

EXGBoost.plot_tree(
  booster,
  rankdir: :lr,
  index: 4,
  depth: 3,
  space_between: [nodes: 200]
)

So if you want to add the space between while not changing the size of the nodes you might need to manually adjust the canvas size:

EXGBoost.plot_tree(
  booster,
  rankdir: :lr,
  index: 4,
  depth: 3,
  space_between: [nodes: 200],
  height: 800
)
EXGBoost.plot_tree(
  booster,
  rankdir: :lr,
  index: 4,
  depth: 3,
  space_between: [levels: 200]
)

Mark Options

The options controlling the appearance of individual marks all conform to a similar API. You can refer to the options and pre-defined defaults for a subset of the allowed options, but you can also pass other options so long as they are allowed by the Vega Mark spec (as defined here)

EXGBoost.plot_tree(
  booster,
  rankdir: :bt,
  index: 4,
  depth: 3,
  space_between: [levels: 200],
  yes: [
    text: [font_size: 18, fill: :teal]
  ],
  no: [
    text: [font_size: 20]
  ],
  node_width: 100
)

Most marks accept an :opacity option that you can use to effectively hide the mark:

EXGBoost.plot_tree(
  booster,
  rankdir: :lr,
  index: 4,
  depth: 3,
  splits: [
    text: [opacity: 0],
    rect: [opacity: 0],
    children: [opacity: 1]
  ]
)

And text marks accept normal text options such as :fill, :font_size, and :font:

EXGBoost.plot_tree(
  booster,
  node_width: 250,
  splits: [
    text: [font: "Helvetica Neue", font_size: 20, fill: "orange"]
  ],
  space_between: [levels: 20]
)

Styles

There are a set of provided pre-configured settings for the top-level API that you may optionally use. You can refer to the EXGBoost.Plottings.Styles docs to see a gallery of each style in action. You can specify a style with the :style option in EXGBoost.plot_tree/2.

You can still specify custom settings along with using a style. Most styles only specify a subset of the total possible settings, but you are free to specify any other allowed keys and they will be merged with the style. Any options passed explicitly to the option does take precedence over the style options.

For example, let's look at the :solarized_dark style:

EXGBoost.Plotting.solarized_dark() |> Keyword.take([:background, :height]) |> IO.inspect()
EXGBoost.plot_tree(booster, style: :solarized_dark)

You can see that it defines a background color of #002b36 but does not restrict what the height must be.

EXGBoost.plot_tree(booster, style: :solarized_dark, background: "white", height: 200)

We specified both :background and :height here, and the background specified in the option supercedes the one from the style.

You can also always get the style specification as a Keyword which can be passed to EXGBoost.plot_tree/2 manually, making any needed changes yourself, like so:

custom_style = EXGBoost.Plotting.solarized_dark() |> Keyword.put(:background, "white")
EXGBoost.plot_tree(booster, style: custom_style)

You can also programatically check which styles are available:

EXGBoost.Plotting.get_styles()

Configuration

You can also set defaults for the top-level API using an Application configuration for EXGBoost under the :plotting key. Since the defaults are collected from your configuration file at compile-time, anything you set during runtime, even if you set it to the Application environment, will not be registered as defaults.

For example, if you just want to change the default pre-configured style you can do:

Mix.install([
  {:exgboost, path: Path.join(__DIR__, ".."), env: :dev},
],
  config:
  [
    exgboost: [
      plotting: [
          style: :solarized_dark,
        ]]
      ],
  lockfile: :exgboost)

You can also make one-off changes to any of the settings with this method. In effect, this turns into a default custom style. Just make sure to set style: nil to ensure that the style option doesn't supercede any of your settings. Here's an example of that:

  default_style =
  [
    style: nil,
    background: "#3f3f3f",
    leaves: [
      # Foreground
      text: [fill: "#dcdccc", font_size: 12, font_style: "normal", font_weight: "normal"],
      # Comment
      rect: [fill: "#7f9f7f", stroke: "#7f9f7f"]
    ],
    splits: [
      # Foreground
      text: [fill: "#dcdccc", font_size: 12, font_style: "normal", font_weight: "bold"],
      # Comment
      rect: [fill: "#7f9f7f", stroke: "#7f9f7f"],
      # Selection
      children: [fill: "#2b2b2b", stroke: "#2b2b2b"]
    ],
    yes: [
      # Green
      text: [fill: "#7f9f7f"],
      # Selection
      path: [stroke: "#2b2b2b"]
    ],
    no: [
      # Red
      text: [fill: "#cc9393"],
      # Selection
      path: [stroke: "#2b2b2b"]
    ]
  ]

Mix.install([
  {:exgboost, path: Path.join(__DIR__, ".."), env: :dev},
],
config:
  [
    exgboost: [
      plotting: default_style,
    ]
  ]
)

NOTE: When you specify a parameter in the configuration, it is merged with the defaults which is different from runtime behavior.

At any point, you can check what your default settings are by using EXGBoost.Plotting.get_defaults/0

EXGBoost.Plotting.get_defaults()

Low-Level API

If you find yourself needing more granular control over your plots, you can reach towards the EXGBoost.Plotting module. This module houses the EXGBoost.Plotting.plot/2 function, which is what is used under the hood from the EXGBoost.plot_tree/2 top-level API. This module also has the get_data_spec/2 function, as well as the to_tabular/1 function, both of which can be used to specify your own Vega specification. Lastly, the module also houses all of the pre-configured styles, which are 0-arity functions which output the Keywords containing their respective style's options that can be passed to the plotting APIs.

Let's briefly go over the to_tabular/1 and get_data_spec/2 functions:

The to_tabular/1 function is used to convert a Booster, which is formatted as a tree structure, to a tabular format which can be ingested specifically by the Vega Stratify transform. It returns a list of "nodes", which are just Maps with info about each node in the tree.

EXGBoost.Plotting.to_tabular(booster) |> hd

You can use this function if you want to have complete control over the visualization, and just want a bit of a head start with respect to data transformation for converting the Booster into a more digestible format.

The get_data_source/2 function is used if you want to use the provided Vega data specification. This is for those who want to only focus on implementing your own Vega Marks, and want to leverage the data transformation pipeline that powers the top-level API.

The data transformation used is the following pipeline:

to_tabular/1 -> Filter (by tree index) -> Stratify -> Tree

EXGBoost.Plotting.get_data_spec(booster, rankdir: :bt)

The Vega fields which are not included with get_data_spec/2 and are included in plot/2 are:

You can make a completely valid plot using only the Data from get_data_specs/2 and adding the marks you need.