Tooltip and grouping information?

[OriolAbril]

We might want to provide tooltip and grouping information somehow as part of the common plotting interface. Even if some backends can't include that into the plot, it would be a great win for those that can.

[OriolAbril]

Here is an example of getting an interactive legend working with plot_dist. As we can pass arbitrary elements as aesthetics and as plot_kwargs everything is feasible already, but we might want to set that behaviour as the default (maybe only when there are multiple models for example, with docs on doing that also for plot_dist with chains as different colors...).

Note: to make that a default the legendgroup argument should become part of the common interface (with this or another name) and be ignored for backends where it makes no sense.

Screencast.de.9-7-24.22.32.03.webm

Code to reproduce

At the time of writing, this requires the plotly_poc branch

import numpy as np
import arviz_plots as azp
from arviz_base import load_arviz_data

import plotly.io as pio

pio.renderers.default = "browser"

c = load_arviz_data("centered_eight")
n = load_arviz_data("non_centered_eight")

pc = azp.plot_dist(
    {"centered": c, "non_centered": n},
    backend="plotly",
    pc_kwargs={"aes": {"legendgroup": ["model"]}, "legendgroup": ["centered", "non_centered"]},
    aes_map={
        "credible_interval": {"color", "y", "legendgroup"},
        "point_estimate": {"color", "y", "legendgroup"},
    },
    plot_kwargs={
        "kde": {"showlegend": False},
        "credible_interval": {"showlegend": False},
        "point_estimate": {"showlegend": False},
        "point_estimate_text": {"showlegend": False},
    }
)

# generate interactive legend
colors = pc.aes["mu"]["color"].values
plot = pc.viz["chart"].item()
obj_centered = pc.viz["mu"]["kde"].sel(model="centered").item()
obj_non_centered = pc.viz["mu"]["kde"].sel(model="non_centered").item()
plot.update_traces(
    selector=lambda trace: (trace.y.shape == obj_centered.y.shape) and np.allclose(trace.y, obj_centered.y),
    name="centered",
    legendgroup="centered",
    showlegend=True,
)
plot.update_traces(
    selector=lambda trace: (trace.y.shape == obj_non_centered.y.shape) and np.allclose(trace.y, obj_non_centered.y),
    name="non centered",
    legendgroup="non_centered",
    showlegend=True,
)
# ------
pc.show()

or with plot_forest:

Screencast.de.9-7-24.22.30.39.webm

Code to reproduce

At the time of writing, this requires the plotly_poc branch

import numpy as np
import arviz_plots as azp
from arviz_base import load_arviz_data

import plotly.io as pio

pio.renderers.default = "browser"

c = load_arviz_data("centered_eight")
n = load_arviz_data("non_centered_eight")
pc = azp.plot_forest(
    {"centered": c, "non_centered": n},
    backend="plotly",
    pc_kwargs={"aes": {"legendgroup": ["model"]}, "legendgroup": ["centered", "non_centered"]},
    plot_kwargs={
        "trunk": {"showlegend": False},
        "twig": {"showlegend": False},
        "point_estimate": {"showlegend": False},
        "labels": {"showlegend": False},
    }
)

# generate interactive legend
colors = pc.aes["mu"]["color"].values
plot = pc.viz["plot"].sel(column="forest").item()
obj_centered = pc.viz["mu"]["point_estimate"].sel(chain=0, model="centered").item()
obj_non_centered = pc.viz["mu"]["point_estimate"].sel(chain=0, model="non_centered").item()
plot.update_traces(
    selector=lambda trace: np.allclose(trace.y, obj_centered.y) & np.allclose(trace.x, obj_centered.x),
    name="centered",
    mode="lines+markers",
    legendgroup="centered",
    showlegend=True,
    line={"color": colors[0]}
)
plot.update_traces(
    selector=lambda trace: np.allclose(trace.y, obj_non_centered.y) & np.allclose(trace.x, obj_non_centered.x),
    name="non centered",
    mode="lines+markers",
    legendgroup="non_centered",
    showlegend=True,
    line={"color": colors[1]}
)
# ------
pc.show()

[aloctavodia]

At least some of this functionality is supported by matplotlib. See here https://matplotlib.org/stable/gallery/event_handling/legend_picking.html

Do we want to support interactivity for matplotlib too?