Hendrik/xgboost
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Optional normalization for learning to rank. (#10094)

Browse Source
This commit is contained in:
Jiaming Yuan
2024-03-08 12:41:21 +08:00
committed by GitHub
parent bc516198dc
commit e14c3b9325
8 changed files with 44 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
doc/parameter.rst
View File

@@ -500,7 +500,11 @@ These are parameters specific to learning to rank task. See :doc:`Learning to Ra
It specifies the number of pairs sampled for each document when pair method is ``mean``, or the truncation level for queries when the pair method is ``topk``. For example, to train with ``ndcg@6``, set ``lambdarank_num_pair_per_sample`` to :math:`6` and ``lambdarank_pair_method`` to ``topk``.
* ``lambdarank_unbiased`` [default = ``false``]
* ``lambdarank_normalization`` [default = ``true``]
Whether to normalize the leaf value by lambda gradient. This can sometimes stagnate the training progress.
* ``lambdarank_unbiased`` [default = ``false``]
Specify whether do we need to debias input click data.

5
doc/tutorials/learning_to_rank.rst
View File

@@ -48,7 +48,7 @@ Notice that the samples are sorted based on their query index in a non-decreasin
import xgboost as xgb
# Make a synthetic ranking dataset for demonstration
seed = 1994
seed = 1994
X, y = make_classification(random_state=seed)
rng = np.random.default_rng(seed)
n_query_groups = 3
@@ -146,7 +146,8 @@ The consideration of effective pairs also applies to the choice of pair method (
When using the mean strategy for generating pairs, where the target metric (like ``NDCG``) is computed over the whole query list, users can specify how many pairs should be generated per each document, by setting the ``lambdarank_num_pair_per_sample``. XGBoost will randomly sample ``lambdarank_num_pair_per_sample`` pairs for each element in the query group (:math:`|pairs| = |query| \times num\_pairsample`). Often, setting it to 1 can produce reasonable results. In cases where performance is inadequate due to insufficient number of effective pairs being generated, set ``lambdarank_num_pair_per_sample`` to a higher value. As more document pairs are generated, more effective pairs will be generated as well.
On the other hand, if you are prioritizing the top :math:`k` documents, the ``lambdarank_num_pair_per_sample`` should be set slightly higher than :math:`k` (with a few more documents) to obtain a good training result.
On the other hand, if you are prioritizing the top :math:`k` documents, the ``lambdarank_num_pair_per_sample`` should be set slightly higher than :math:`k` (with a few more documents) to obtain a good training result. Lastly, XGBoost employs additional regularization for learning to rank objectives, which can be disabled by setting the ``lambdarank_normalization`` to ``False``.
**Summary** If you have large amount of training data: