rfvis 0.3.2

A tool for visualizing the structure and performance of Random Forests

RFVis

A tool for visualizing the structure and performance of Random Forests (and other ensemble methods based on decision trees).

RFVis offers a Command Line API and a Python API which works on a sklearn.ensemble.RandomForestClassifier.

Getting Started

Install and update RFVis via pip:

$ pip install rfvis

This will allow you interactively visualize a fitted Random Forest (RF) in your browser. To directly generate SVG files from your model you also need to install Node.js, see Command Line Interface for more information.

Command Line API

RFVis offers a command line tool to either generate SVG files directly from your input data (rfvis cli <data>) or to spin up a web-based GUI for a more interactive analysis (rfvis gui <data>).

To see all available commands run:

$ rfvis --help
Usage: rfvis [OPTIONS] COMMAND [ARGS]...

  A tool for visualizing the structure and performance of Random Forests

Options:
  --version  Show the version and exit.
  --help     Show this message and exit.

Commands:
  cli  Command line interface to generate SVGs.
  gui  Web-based graphical user interface.

Graphical User Interface

To interactively analyze your forest with the web-based GUI run:

$ rfvis gui /path/to/data
 * Running on http://127.0.0.1:8080/ (Press CTRL+C to quit)

You can now open up your browser at http://localhost:8080 to see something like this:

Command Line Interface

To use the Command Line Interface (CLI) you need to have Node.js v8+ installed on your system. This is a technical limitation due to the fact that the rendering is written in Javascript. You do not need to install any other package though, the CLI integrates into the command line tool you already installed via pip:

$ rfvis cli /path/to/data
>> Exported "/dev/random-forest-visualization/tree-0.svg"
>> Exported "/dev/random-forest-visualization/tree-1.svg"
>> Exported "/dev/random-forest-visualization/tree-2.svg"
>> Exported "/dev/random-forest-visualization/tree-3.svg"
...

Get a full list of available options with --help:

$ rfvis cli --help
Usage: rfvis cli [OPTIONS] FOREST_JSON

  Web-based graphical user interface.

  As Python is unable to render React components, we make a subprocess call to a small
  Node.js application which will do the rendering and also store the created SVG
  files. This command requires that Node.js is installed on your system!

  FOREST_JSON: Path to the JSON file that contains the forest's data.

Options:
  -o, --out PATH                  Output path of the SVG files.  [default: (current
                                  working directory)]
  -w, --width INTEGER             Width of the SVG.  [default: 800]
  -h, --height INTEGER            Height of the SVG.  [default: 800]
  --trunk-length INTEGER          Length of the trunk which influences the overall tree
                                  size.  [default: 100]
  --display-depth INTEGER         Maximum depth of the tree rendering. Cut of leaves are
                                  visualized as pie chart consolidation nodes.
  --branch-color [Impurity]       Coloring of the branches.  [default: Impurity]
  --leaf-color [Impurity|Best Class]
                                  Coloring of the leaves.  [default: Impurity]
  --help                          Show this message and exit.

Input Data

The data for the Command Line API must be available on your filesystem as a JSON file for the forest and additionally one CSV file per tree. Both data formats will be extended with properties in the future, this is just the minimal set.

You can find a working example under examples/PolSAR.

Forest JSON

The main forest.json holds all information about the ensemble model:

• name (string): Name of your forest, will be displayed in the GUI
• error (float): The error (e.g. the out-of-bag or validation error) of the entire ensemble model, will be displayed in the GUI
• n_samples (int): Number of samples the model was trained on
• correlationMatrix (float[][]): Correlation between the single trees within the model. Has dimensions NxN where N is the number of trees. This will be used to compute the forest map.
• classes: The output classes
  • name (string): Name of the class
  • color (int, int, int): RGB values in the range of 0-255 which determine the color of the class in the visualization
• trees: The trees in the forest
  • error (float): The error (again could be either the out-of-bag or validation error) of the single tree
  • data (string): Relative path to the CSV file containing the tree data

Tree CSV

For each tree specified in the forest.json RFVis expects a CSV file where one entry represents one node in the tree. An entry has the following format:

• id (int): ID of the node
• depth (int) Depth of the node in the tree (starting at 0)
• n_node_samples (int): Number of training samples reaching the node
• impurity (float): Impurity of the node (0-1)
• value (int[]): Class distribution within the node, i.e. every entry represents the amount of samples within the node that respond to a specific class. The index corresponds to the indices in forest.classes.

Python API

RFVis also offers a Python API which works directly on a scikit-learn RandomForestClassifier. You can find a working example under examples/scikit_learn.py.

The function rfvis.gui() visualizes a fitted RandomForestClassifier in a web based graphical user interface. The server runs in a separate process and is available at http://localhost:<port>.

gui(model, data=None, target=None, name=None, class_names=None, class_colors=None, port=8080)

Args:

• model (sklearn.ensemble.RandomForestClassifier): The model to visualize.
• data (array-like, shape=(n_samples, n_features)): The training input samples that were used to fit the model. Used to compute the out-of-bag error and correlation of the individual trees. If not provided, the forest view will have no significance.
• target (array-like, shape=n_samples): The target values (class labels) that were used to fit the model. Used to compute the out-of-bag error and correlation of the individual trees. If not provided, the forest view will have no significance.
• name (str): Optional name of the model which will be displayed in the frontend.
• class_names (List[str]): Optional list of names of the target classes
• class_colors (List[str]): Optional list of browser interpretable colors for the target classes. See https://developer.mozilla.org/en-US/docs/Web/CSS/color_value.
• port (int): Port on which the frontend will run on. Defaults to 8080.

Returns:

• process (multiprocessing.Process): Subprocess that runs the server. Can be terminated with process.terminate().

Development

The repository contains a Pipfile for conveniently creating a virtualenv for development. Just install pipenv and run:

$ pipenv install

You can now e.g. start the server on the default port 8080 via:

$ pipenv run rfvis gui <path_to_forest_json>

Note that you need to build the frontend bundle first before you can actually see the application working on http://localhost:8080.

To build the frontend you need Node.js installed. First install all dev-dependencies by running the following from within the ./rfvis/client directory:

$ npm install

Now you can build a production-ready bundle via:

$ npm run build

If you have the Python server running you should now be able to see the application at http://localhost:8080.

For developing on the frontend more conveniently run:

$ npm start

To start a development server with hot reloading at http://localhost:3000.

Publications

If you are using RFVis in your research, please cite the following paper:

• Ronny Hänsch, Philipp Wiesner, Sophie Wendler, and Olaf Hellwich. "Colorful Trees: Visualizing Random Forests for Analysis and Interpretation" In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pp. 294-302. IEEE, 2019.