Feature Request: px.distribution_drift, px.model_disagreement, px.quantile_evolution — ML Monitoring Primitives

[samibahig]

Feature Request: px.distribution_drift, px.model_disagreement, px.quantile_evolution — ML Monitoring Visualization Primitives

Hi Plotly maintainers,

I'd like to propose three new high-level functions for plotly.express that address a real gap in the ML observability and data science space. These are general-purpose visualization primitives — not domain-specific — that fill gaps common in production ML workflows.

I've built working prototypes with full test coverage and a live interactive demo. Happy to take these through the full contribution process.

---

1. px.distribution_drift(reference, current, ...)

What it does: Compares two distributions (e.g. training vs. live inference data) with overlapping normalized histograms and a scalar KL-divergence annotation. When drift exceeds a configurable threshold, the current distribution is highlighted in a warning color.

fig = px.distribution_drift(
    reference,                # array-like: baseline samples
    current,                  # array-like: current samples
    bins=50,
    divergence_threshold=0.1,
    reference_name="Reference",
    current_name="Current",
    title=None,
    template=None,
)

Gap it fills: px.histogram with barmode="overlay" gets you close, but there's no built-in divergence scoring, threshold annotation, or drift-aware color logic. This is a one-liner for a very common ML monitoring pattern.

---

2. px.model_disagreement(x, y, predictions, ...)

What it does: Scatter plot of samples in a 2-D reduced feature space (UMAP/t-SNE/PCA), colored by ensemble variance. Samples above a disagreement threshold are fully opaque; others are dimmed. Includes a marginal histogram of variance scores.

fig = px.model_disagreement(
    x,                        # dim 1 of reduced space (per sample)
    y,                        # dim 2 of reduced space
    predictions,              # shape (n_samples, n_models) — ensemble preds
    threshold=0.05,
    colorscale="Viridis",
    title=None,
    template=None,
)

Gap it fills: px.scatter with color= handles the spatial encoding, but computing ensemble variance, dimming low-uncertainty samples, and combining with a marginal histogram requires significant boilerplate. This surfaces a critical active-learning and model-audit pattern as a single call.

---

3. px.quantile_evolution(timestamps, p50, *, p10, p90, p25, p75, ...)

What it does: Layered ribbon/band chart tracking P10–P90 and P25–P75 bands with P50 as a bold center line, over time or cohorts. Automatically annotates timestamps where the spread exceeds a volatility threshold.

fig = px.quantile_evolution(
    timestamps,               # x-axis: ISO strings, labels, or numbers
    p50=median_values,        # required: median (center line)
    p10=p10_values,           # optional: outer lower band
    p90=p90_values,           # optional: outer upper band
    p25=p25_values,           # optional: IQR lower bound
    p75=p75_values,           # optional: IQR upper bound
    mean=mean_values,
    show_mean=False,
    volatility_multiplier=1.5,
    title=None,
    template=None,
)

Gap it fills: Building fill-between ribbon plots requires chaining 5+ go.Scatter traces with careful fill="tonexty" sequencing. This is error-prone and undiscoverable. A single px call makes this pattern accessible to the full data science audience.

---

Why these belong in plotly.express

• Each returns a go.Figure — composable and customizable after the fact
• They accept lists, numpy arrays, and pandas Series via a thin _to_list wrapper
• They use only existing trace types (go.Histogram, go.Scatter) — no new trace types needed
• They follow the existing px design philosophy: one call, sensible defaults, everything overridable

---

Working prototype

• GitHub: https://github.com/samibahig/plotlyml
• Live demo (Streamlit): https://huggingface.co/spaces/samibahig-md/plotlyml-demo

Zero dependencies beyond plotly itself. 12-test suite passes on Python 3.9–3.12.

I'm ready to adapt the implementation to Plotly's internal conventions, add API docs, write tests in the tests/test_core/test_px/ format, and address any design feedback on the signatures.

Thank you for considering it!