dash-upset

Interactive UpSet plots for Plotly Dash.

Documentation and live explorer

What is an UpSet plot?

An UpSet plot visualizes the intersections of many sets. Venn and Euler diagrams become unreadable past three or four sets; UpSet replaces the overlapping circles with:

• a matrix whose columns are sets and whose rows are intersections (filled, connected dots show which sets participate in each intersection),
• set-size bars giving the cardinality of each individual set, and
• intersection-size bars giving the size of each intersection.

This scales to dozens of sets and makes the large intersections obvious at a glance. dash-upset brings it to Dash as a reusable, themeable, callback-friendly component.

Why dash-upset

• Pure Python, MIT-licensed. All UpSet logic (counting modes, sorting, filtering, deviation, theming) composes a Plotly figure in Python. No heavy JavaScript stack, and it renders in notebooks.
• Drop-in Dash component. UpSet renders that figure and turns clicks into ordinary component properties, so callbacks use the standard Input(id, property). The compiled React layer ships prebuilt, so installing needs no Node toolchain.
• One input, two ways to render. The same data drives a live UpSet component or a static create_upset figure for notebooks and publication export (PNG / SVG / PDF).

Installation

pip install dash-upset

conda-forge: A recipe has been prepared and is pending submission to conda-forge/staged-recipes. Once the feedstock is live, conda install -c conda-forge dash-upset will work. Until then, install from PyPI with pip (or pixi add --pypi dash-upset in a pixi project).

Quick start

Drop the UpSet component into a Dash layout with a dataframe of boolean indicator columns (one per set). Clicks surface as component properties your callbacks read the standard Dash way:

import pandas as pd
from dash import Dash, Input, Output, callback, html
from dash_upset import UpSet

# One row per misclassified test example; 1 = that model got it wrong.
# Overlaps are the shared hard cases; singletons are each model's blind spots.
df = pd.DataFrame(
    {
        "ResNet": [1, 1, 0, 1, 0, 1],
        "ViT": [1, 1, 1, 0, 0, 1],
        "XGBoost": [0, 1, 1, 1, 1, 0],
    }
)

app = Dash(__name__)
app.layout = html.Div(
    [
        UpSet(id="errors", data=df, sets=["ResNet", "ViT", "XGBoost"]),
        html.Pre(id="out"),
    ]
)


@callback(Output("out", "children"), Input("errors", "selected_intersection"))
def show(selection):
    # {"label": "ResNet & ViT", "sets": ["ResNet", "ViT"], "size": 2}
    return str(selection)


if __name__ == "__main__":
    app.run(debug=True)

selected_intersection updates when an intersection-size bar or a matrix dot is clicked; selected_sets (a list of set names) updates when a set-size bar is clicked.

Just the figure

For notebooks, scripts, or static export, create_upset takes the same input and returns a plain plotly.graph_objects.Figure:

from dash_upset import create_upset, from_counts

fig = create_upset(
    from_counts({
        "Action": 320, "Comedy": 290, "Drama": 410,
        "Action&Comedy": 84, "Action&Drama": 120, "Comedy&Drama": 96,
        "Action&Comedy&Drama": 40,
    }),
    title="Movie genres",
)
fig.show()

Element-level data uses the familiar upsetplot conventions:

from dash_upset import from_contents, from_indicators, from_memberships

from_memberships([("A",), ("A", "B"), ()])          # per-element set names
from_contents({"A": ["x", "y"], "B": ["y", "z"]})   # per-set element ids
from_indicators(boolean_dataframe)                  # rows = elements, columns = sets

from_indicators is dataframe-agnostic via narwhals: pandas, Polars, PyArrow, cuDF, and Modin frames (or a plain dict of boolean columns) all work, and dash-upset itself depends on none of those libraries.

Sorting and display are controlled per plot, e.g. UpSet(data=df, sets=[...], sort_by="degree", sort_sets_by="name", theme="dark"); create_upset accepts the same keywords. The full argument reference lives at https://phylatech.github.io/dash-upset/reference.html.

Contributing

Development setup, the docs workflow, and the commit conventions that drive releases are in CONTRIBUTING.md.

Prior art and credits

• UpSet and the original research by Lex, Gehlenborg, et al. define the technique.
• UpSet 2.0 by the Visualization Design Lab (BSD-3-Clause) is the technique authors' interactive reimplementation.
• UpSet.js by Samuel Gratzl is an interactive JS implementation (AGPLv3 / commercial).
• upsetplot by Joel Nothman (BSD-3-Clause) is the established matplotlib-based Python package; dash-upset mirrors its familiar data-input conventions.

License

MIT © Evan Roy Rees