[python] [R-package] refine the parameters for Dataset

[guolinke]

to fix #2590

• remove the pre-filter by min_data

[guolinke]

@StrikerRUS could you help to add the tests, to ensure the warnings are shown as expected?

[StrikerRUS]

@guolinke Sure, I'll try!

[guolinke]

@StrikerRUS I think we can move the discussion to the main thread. it is hard to notice the new conversion in the above sub-thread.

[StrikerRUS]

@guolinke

unfortunately, It is not trivial. As we don't know the filename passed in c_api is training or k-th valid data. It is hard to retrieve the corresponding init score filename. The easier way in python/R is read the init_score_file by user, and set init_score field in dataset.

another solution is to force the init_score_filename to concat(data_filename, ".init"), and remove all initscore filename parameters.

I think removing filename params is good idea! It will decrease overall number of LightGBM params which is good change for newcomers who struggling with learning a huge number of settings. Also, it will fix inconsistency in behavior when for language packages default filename is respected but it cannot be changed by those params. As we've already removed some sparse-related params in #2699 and minor (or major) version bump is required, it is OK in terms of backward compatibility.

[guolinke]

yeah, maybe do it in a new PR?

[StrikerRUS]

@guolinke

yeah, maybe do it in a new PR?

Sure, as you wish!

[guolinke]

ping @jameslamb for R tests.
@StrikerRUS any other issues need to be addressed?

[StrikerRUS]

@guolinke

any other issues need to be addressed?

As we agreed on making the breakage in docs into Dataset and Booster params, and removing init_score params in a separate PRs, I think this PR is ready except R side (aliases and tests).

[guolinke]

ping @jameslamb for R's part. Maybe R test could be in a new PR?

[jameslamb]

ping @jameslamb for R's part. Maybe R test could be in a new PR?

just added #2767 . Let me know if you think anything is missing!