How do I share a SHAP force plot?

`shap.save_html('shap.html', shap.force_plot(...))` writes a standalone interactive page, then publish that file — one curl, or drag it into the app. The plot's JavaScript runs at the link, so hovering and reordering still work.

Why doesn't the force plot render when I send the file?

Force plots are JavaScript-driven — that's why notebooks need `shap.initjs()`. Saved with `save_html` the runtime is included, but a recipient opening it from a mail attachment hits the usual local-file problems; a URL avoids all of them.

What about the matplotlib SHAP plots?

Summary, beeswarm, waterfall and dependence plots render to images. Save them as PNGs, write a small HTML page around them with your interpretation, and publish that — one report that carries the explanation and the evidence together.

The people who assess the model and don't run notebooks: risk, compliance, the product owner, the domain expert who knows why a feature is suspicious. They need a URL, not a repo.

Can they comment on a specific feature?

Yes. Threads anchor to the rendered content, so 'this feature is a leak — it's populated after the label' lands on that feature and stays on the record through the next model version.

Share a SHAP plot

Model explanation has an audience problem. The person who most needs to see which features drove a prediction — a risk reviewer, a clinician, a product owner, the domain expert who will spot the leaked feature — is the person least likely to open a notebook.

So the explanation gets screenshotted into a deck, and the interactive part, which is the part that answers follow-up questions, is gone.

Publish the interactive plot

import shap

explainer = shap.TreeExplainer(model)
shap_values = explainer.shap_values(X_test)

shap.save_html("shap.html", shap.force_plot(explainer.expected_value, shap_values, X_test))

curl -X POST https://commareports.com/api/v1/reports \
  -H "Authorization: Bearer $COMMA_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d "$(jq -n --rawfile html shap.html \
        '{title: "Churn model v4 — SHAP", html: $html, visibility: "registered"}')"

Or drag shap.html into the app. The plot's JavaScript runs inside a sandboxed iframe (allow-scripts, no allow-same-origin), so hovering, reordering and the sample selector work at the link — see interactive HTML reports.

The static plots, with the argument around them

Summary, beeswarm, waterfall and dependence plots are matplotlib figures. Save them and wrap them in the page you'd otherwise have written in Slack:

import base64, io, matplotlib.pyplot as plt

shap.summary_plot(shap_values, X_test, show=False)
buf = io.BytesIO(); plt.savefig(buf, format="png", bbox_inches="tight", dpi=144)
img = base64.b64encode(buf.getvalue()).decode()

html = f"""
<h1>Churn model v4 — what drives the score</h1>
<p><code>days_since_last_login</code> dominates, as expected.
<code>support_tickets_30d</code> is third, which is new in v4 and worth a look.</p>
<img src="data:image/png;base64,{img}" alt="SHAP summary plot">
"""

One report holding the plots and the interpretation is a better review artifact than either alone.

Why the link matters for model review

Anchored threads on the feature someone distrusts — the leak, the proxy variable, the one that shouldn't be predictive. See commenting on HTML.
Revisions per model version, so v5's explanation sits at the same URL as v4's and the two can be diffed.
Access control — private, team, domain-gated or named reviewers, which matters when the features are customer attributes. See the sharing model.
A record of the sign-off, which is what a model-risk process asks for and a notebook cell can't hold.

Limits

Entry HTML: 5 MB. A force plot over thousands of rows inlines all of them — sample first (X_test.sample(500)), which is also the more readable plot.
60 requests/minute per token.

Try it

Comma is free — unlimited reports, unlimited commenters, unlimited revision history.

Publish a model report →