We have trained the model using the LambdaMART algorithm, where the final predicted score is the weighted sum of the prediction of each regression tree in the ensemble.
So the output of the model is an arbitrary score that doesn’t reflect the relevance labels (test labels) we use in training/test sets but it is just a measure of how relevant the document was for that query and therefore an indicator of relative ordering between documents.
Consider an example where for a given query you have 4 documents:
You use the predicted scores to sort descending the documents in order to obtain the ranking predicted by the model:
Then the NDCG is computed using the relevance labels.
Hi @cuongnc220592,
We have trained the model using the LambdaMART algorithm, where the final predicted score is the weighted sum of the prediction of each regression tree in the ensemble.
So the output of the model is an arbitrary score that doesn’t reflect the relevance labels (test labels) we use in training/test sets but it is just a measure of how relevant the document was for that query and therefore an indicator of relative ordering between documents.
Consider an example where for a given query you have 4 documents:
You use the predicted scores to sort descending the documents in order to obtain the ranking predicted by the model:
Then the NDCG is computed using the relevance labels.