Xgboost: Understanding tree structure for multiclass classification

[sachinruk]

In my problem I have 3 classes to predict and I'm trying to understand how the xgboost tree relates to the predicted probabilities:

model2=xgboost(data = as.matrix(train_set), label = output_vector, max.depth = 3,
               eta = 1, nround = 50,objective = "multi:softprob",num_class=3,
               verbose = FALSE)
names <- dimnames(train_set)[[2]][-c_set]
trees=xgb.model.dt.tree(names,model = model2)

The first few lines of the tree is as below:

    ID      Feature Split Yes  No Missing   Quality     Cover Tree Yes.Feature Yes.Cover Yes.Quality
1: 0-0 location1100   0.5 0-1 0-2     0-1 26.142200 2666.6700    0 location995   2651.56   22.050700
2: 0-1  location995   0.5 0-3 0-4     0-3 22.050700 2651.5600    0 location962   2637.33   14.790700
3: 0-2         Leaf    NA  NA  NA      NA -0.579310   15.1111    0          NA        NA          NA
4: 0-3  location962   0.5 0-5 0-6     0-5 14.790700 2637.3300    0        Leaf   2620.44    0.723138
5: 0-4         Leaf    NA  NA  NA      NA -0.503650   14.2222    0          NA        NA          NA
6: 0-5         Leaf    NA  NA  NA      NA  0.723138 2620.4400    0          NA        NA          NA
   No.Feature No.Cover No.Quality
1:       Leaf  15.1111  -0.579310
2:       Leaf  14.2222  -0.503650
3:         NA       NA         NA
4:       Leaf  16.8889  -0.204969
5:         NA       NA         NA
6:         NA       NA         NA

how would this relate to the classes as there is only a "yes" or "no" split. Even the leafs do not mention a class that they belong to nor a probability associated. I guess in some ways what I'm asking is how would I manually calculate the probabilities for a test case given the tree structure (in my case in the variables trees).

[Far0n]

A tree always seperates only two classes. Hence, if you have 3 classes, xgb grows 3 trees per boosting round (class 1 vs rest, class 2 vs rest & class 3 vs rest).