Remove unbiased LambdaRank

[RAMitchell]

Summary

• remove the unbiased LambdaRank implementation and its learned position-bias state from CPU and CUDA paths
• retain lambdarank_unbiased as a compatibility parameter that warns once and falls back to standard LambdaRank
• remove the unbiased demo and parameter-generation coverage, and update Python, R, and tutorial documentation
• replace unbiased behavior tests with fallback and legacy model-config loading coverage

Motivation

Unbiased LambdaRank makes the objective stateful by learning and serializing position-bias vectors. This complicates model I/O, checkpoint/resume behavior, cache dimensions, and parameter changes. Removing the implementation simplifies LambdaRank ahead of separating its CPU and GPU kernels while preserving compatibility for existing configurations.

Closes #12302.

User impact

Setting lambdarank_unbiased=true no longer enables unbiased LambdaRank. XGBoost emits a warning and trains with standard LambdaRank. Existing saved configurations containing the parameter or the old bias vectors continue to load.

Validation

• CPU testxgboost build
• ./build/testxgboost --gtest_filter="LambdaRank.*:NDCGCache.*" (9 tests passed)
• affected CUDA objective and test translation units compiled
• pre-commit hooks and repository-rooted C++ header lint
• git diff --check

[Copilot]

Pull request overview

Removes stateful unbiased LambdaRank while preserving lambdarank_unbiased as a warning-and-fallback compatibility option.

Changes:

• Removes CPU/CUDA position-bias state and kernels.
• Adds fallback and legacy-config tests.
• Updates demos and Python/R documentation.
• XGBoost4J-Spark parameters remain outdated and require changes.

Reviewed changes

Copilot reviewed 15 out of 16 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
src/common/ranking_utils.h	Removes bias parameters and cache state.
src/objective/lambdarank_obj.h	Simplifies shared gradient interfaces.
src/objective/lambdarank_obj.cuh	Removes CUDA bias inputs.
src/objective/lambdarank_obj.cc	Implements warning and standard-ranking fallback.
src/objective/lambdarank_obj.cu	Removes CUDA unbiased calculations.
tests/cpp/objective/test_lambdarank_obj.h	Updates test declarations.
tests/cpp/objective/test_lambdarank_obj.cc	Tests fallback and legacy loading.
tests/cpp/objective/test_lambdarank_obj.cu	Updates CUDA tests.
tests/cpp/plugin/test_sycl_lambdarank_obj.cc	Updates SYCL fallback test.
tests/python/test_ranking.py	Removes unbiased behavior test.
python-package/xgboost/testing/params.py	Removes unbiased generated parameters.
R-package/R/xgb.train.R	Updates R parameter API documentation.
R-package/man/xgb.params.Rd	Regenerates R documentation.
doc/parameter.rst	Documents compatibility fallback.
doc/tutorials/learning_to_rank.rst	Documents feature removal and alternative.
demo/guide-python/learning_to_rank.py	Removes click-debiasing demo.

Files not reviewed (1)

• R-package/man/xgb.params.Rd: Generated file

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[trivialfis]

I'm not quite sure about this yet. The state seems quite trivial to fix.

[RAMitchell]

There are a couple more issues:

• What happens if the dmatrix changes during training?
• What happens when you slice the model?

Stateful objectives break everything.

[trivialfis]

Both issues are about training continuation:

• Training continuation with a different DMatrix.
• Slice the model then continue training?

Does it make sense to do such things?