Enable use of constructed Dataset in predict() methods

[jameslamb]

Summary

It should be possible to call predict() on a constructed Dataset and get raw predictions.

Note that it is currently possible to get an evaluation (value of metrics) on a constructed dataset in LightGBM's public API. See #4034 (comment).

Motivation

Getting raw predictions can be useful for purposes other than evaluation of how a model performs on some set of metrics.

For example, in LightGBM the predict() method can be used for purposes like analyzing which leaf nodes each sample matches in each tree (pred_leaf). Without the ability to call predict() on a constructed Dataset, doing such analysis requires loading the raw data in memory.

Description

N/A

References

• [R-package] predict() breaks when using a Dataset stored in a file #4034
• Support specifying number of iterations in dataset evaluation #4210

[jameslamb]

I opened this feature request in response to #4034 (comment).

Added it to #2302, per this project's standard practice. If you are interested in contributing this feature, please leave a comment here and maintainers can re-open the issue and answer any questions you have.

tagging @shiyu1994 and @StrikerRUS , in case you want to add any additional details.

[StrikerRUS]

This is duplicate of #1939.

[jameslamb]

This is duplicate of #1939.

oh! Forgot about that one. Since it isn't in #2302 I didn't think we already had a feature request for it.

I'll add a link to that to the existing item for this feature in #2302.