[RFC] Unify appearance of `plot_param_importances` between plotly and matplotlib backends

Motivation

Plotly’s plot_param_importances shows the importance scores explicitly as in the optuna’s docs page on the right next of each bar. On the other hand, mathplotlib’s one does not show the scores as in the optuna’s docs page there. To minimize the difference between them, we would like to see similar plots not depending on the backends.

I think we have two options:

1. Remove importance scores from Plotly’s plot
2. Add importance scores to matplotlib’s plot

The reason behind the first option is that the scores mislead users sometimes, particular when the importance scores are less than 0.01. See https://github.com/optuna/optuna/issues/2868#issue-978621430 as an example. Note that users can modify the plot object returned by plot functions in practice; users can remove them or adjust importance scores. In addition, if optuna internally removes the importance scores, it is hard for users to compute the importance score again due to its stochastic nature (see also related issue https://github.com/optuna/optuna/issues/2757).

Description

Depending on the decision.

Alternatives (optional)

Additional context (optional)

A similar plot by scikit-learn does not provide such values.

[nzw0301]

I think we can close this issue because @xadrianzetx resolved this issue by #3012 and #3073. Thank you so much for improving both plot_param_importances functions!

[nzw0301]

Let me summarise the opinions from dev-members @HideakiImamura @hvy @toshihikoyanase.

• Keep numerical values on the right next to bars. Technically we can add std bar for the current importance score’s algorithms. However, we cannot add an error bar as on sklearn’s example if we introduce a deterministic importance calculation algorithm.
• Add importance scores to matplotlib’s plot to unify the appearance.
• For an importance score is less than 0.01, change the format to <0.01. (further opinions are welcome!)

Now, I suppose that we can resolve this issue by sending two PRs:

• change the format for importance score less than 0.01
• add the same importance score’s text to matplotlib backend as Plotly’s one.