[BUG] It seems that the prediction of objective "multi:softprob" is incorrect and the order of the probability is inconsistent with mllib.

[sonetto19999]

multi:softmax output:

multi:softprob output:

these two figure clearly shows that the prediction of "multi:softprob" is incorrect, the first three rows shold be 0 rathe than 1 (just like the prediction of "multi:softmax")

Meanwhile, from what I recall, the order of probabilities in MLlib is the same as the labels when the labels increase while it is not true in XGBoost

[trivialfis]

cc @wbo4958

[sonetto19999]

anyone help?

[mayer79]

@sonetto19999 You might expect more help with a minimally reproducible example

[wbo4958]

Sorry for delay, I just see this issue. That's terrible. seems the prediction values of softprob is reversed of softmax.

Hi @sonetto19999, were you using xgboost4j-spak or xgboost pyspark?

[sonetto19999]

@wbo4958 currently using xgboost4j-spark_2.12

[wbo4958]

I can repro this issue when setting the objective = multi:softmax or multi:softprob and num_class = 2 for the binary classification.

Hi @trivialfis, is this scenario allowed in xgboost?

[trivialfis]

If it's binary classification, it should be using binary:logistic instead of softmax.