interaction_constraints with more than 10^5 parameters fail.

[palVJ]

Using xgb.train with more than 10^5 parameters and using the parameter interaction_constraints, I get the following error:

Error in xgb.iter.update(bst$handle, dtrain, iteration - 1, obj) :
[09:54:20] ./dmlc-core/include/dmlc/././json.h:788: Check failed: ch == ',' (101 vs. ,) : Error at Line 0, around ^+05,100001,100002,100003,100004,100005,100006,100007,100008,100, JSON array expect ']' or ','. Get 'e' instead

I suspect 100001, 100002 and so on represent feature indices with values more than 100 000.
I tried with feature indices with values less than 100 000, and it worked.

Am I right, and if so, can this be fixed?

[trivialfis]

Em.. there's scientific notation in the indices string. Should be able to fix it.

[palVJ]

Using R and input parameter interaction_constraints does not contain numbers with scientific notation. Seems there is something to do with the source code, but I may be wrong. tor. 2. jan. 2020, 16:28 skrev Jiaming Yuan <notifications@github.com>:

…

Em.. there's scientific notation in the indices string. Should be able to fix it. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#5179?email_source=notifications&email_token=AHUJYHLSN3T4XZA4TN4WZLDQ3YBZ3A5CNFSM4KCDM2N2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEH6TBKI#issuecomment-570241193>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AHUJYHIPMYBH6C6BVGSSEK3Q3YBZ3ANCNFSM4KCDM2NQ> .

[trivialfis]

The parameters are passed as string into XGBoost, a quick workaround right now is to disable R's scientific representation used by paste. Will further investigate a proper solution.