What is the result of XGboost using "rank: pairwise"?

Question

I am using python implementation of XGBoost. One of the goals is rank:pairwise and minimizes pairwise losses ( Documentation ). However, he does not say anything about the volume of output. I see numbers from -10 to 10, but can there really be an inf inf?

Answer 1

good question. You can look into the kaggle competition :

In fact, in the Learning to Rank field, we are trying to predict the relative score for each document for a specific request. That is, it is not a regression problem or a classification problem. Therefore, if a document attached to a request receives a negative forecast indicator, it means and only means that it is relatively less relative to the request when compared with other documents with positive ratings.

Answer 2

If I understand your qns correctly, you mean the output of the prediction function on the model installed using the rank: pairwise.

Prediction gives a predicted variable (yhat).

This is the same for reg: linear / binary: logistic, etc. The only difference is that reg: linear builds the trees in min (rmse (y, yhat)), and rank: the assembly trees fall in max (map (rank (y), rank (yhat))). However, the conclusion is always yhat.

Depending on the values ​​of your dependent variables, the output may be any. But I usually expect that the result will be much smaller compared to the dependent variable. This usually takes place, since there is no need to correspond to extreme data values, the tree just has to create predictors that are large / small in order to rank first / in the group in the group

Answer 3

It gives a predictable ranking score. However, points are only valid for ranking in their own groups. Therefore, we must set up groups for input.

For esay ranking refer to my xgboostExtension project