What's the difference between multiclass categorical crossentropy, mlogloss and multi:softprob?

Question

As far as I understand, an objective is something I'm trying to optimize and an evaluation statistic is something I use to look for overfitting. I stumbled upon 4 losses that seem to be the same, but I'm not quite sure.

1. In XGBoost, one of the objectives for multi class classification is multi:softprob (https://xgboost.readthedocs.io/en/latest/parameter.html#learning-task-parameters). I couldn't find any formulas in the documentation, but apparently this objective returns "predicted probability of each data point belonging to each class", so an "ndata * nclass matrix".
2. One of XGBoost's eval_metrics is mlogloss. The documentation (same link as above) links to sklearn.metrics.log_loss, which is "log loss, aka logistic loss or cross-entropy loss". sklearn's User Guide about log loss provides this formula: $$ L(Y, P) = -\frac1N \sum_i^N \sum_k^K y_{i,k} \log p_{i,k} $$ So apparently, mlogloss and (multiclass categorical) cross-entropy loss are the same.
3. The Otto Group Product Classification Challenge talks about "the multi-class logarithmic loss" and gives the same formula as above, so looks like mlogloss, cross-entropy loss and multi-class logarithmic loss are all the same.

Now, I'm not particularly sure what multi:softprob is. multi:softmax is softmax - the activation function used in neural networks, for example. However, it's not a measure of error, like cross-entropy loss, right? Why is it one of the objectives in XGBoost, then? Unfortunately, there doesn't seem to be any useful information about multi:softprob, except that it's not the same as softmax because softprob outputs a vector of probabilities and softmax - "a class output" (so the ID of a class, I presume?).

---

Am I correct that mlogloss, cross-entropy loss and multi-class logarithmic loss are the same thing? How are they different from multi:softprob? What is multi:softprob anyway?

Answer 1

From the documentation of XGBoost you can see that multi:softmax and multi:softprob are the same objective. The only difference is that multi:softprob also return output vector of ndata * nclass of the classes probabilities.

multi:softmax: set XGBoost to do multiclass classification using the softmax objective, you also need to set num_class(number of classes)

multi:softprob: same as softmax, but output a vector of ndata * nclass, which can be further reshaped to ndata * nclass matrix. The result contains predicted probability of each data point belonging to each class.