Roadmap: feature requests

[hcho3]

This issue is aimed at keeping track of requested features.

Note to developers and contributors: If you plan to actively work to implement a requested feature, re-open the linked issue. Everyone is welcome to work on any of the issues below. For specifics, read the linked issues.

Note to maintainers: All issues with feature requests should be consolidated to this document. Close all new issues with feature requests and create corresponding new entries in the following checklist. (Don't forget to create a link to the closed issue.) This is so that the number of open issues is kept manageable. Also make sure to attach feature-request label when closing feature requests

• Multiple output regression (Multiple output regression #2087)
• Rust binding (Feature request: Add a Rust API #3351)
• Lua binding (Lua Wrapper to XGBoost #1444)
• Save data counts inside XGBoost model ([jvm-package] How to get the record count number of each leaves in a well trained model? #3419)
• Feature interaction constraint ([New Feature] Interaction constraints #3135)
• Clarify behavior for trivial splits (Clarification on Trivial Splits #2914)
• Stratified distributed DMatrix (Stratified distributed DMatrix #2758)
• Dynamic min_child_weight parameter (Feature request: Dynamic min_child_weight parameter #2714)

Distributed learning

• Write a tutorial for distributed learning using Docker containerization ([Feature request] Docker image for distributed version of xgboost #3593)
• Provide a tutorial for using XGBoost with MPI (distributed xgboost mpi examples #3596)

Python specific

• Preserve feature_names when slicing DMatrix with slice() (Python: DMatrix.slice() looses feature_names #3124)
• Pre-built binary wheels for Windows
• Automatically set ntree_limit = best_ntree_limit for predictions in scikit-learn API (suggestion: default to ntree_limit = best_ntree_limit for predictions in scikit-learn API #3053)
• Make prediction at each stage for a given data matrix (Faster way to get predictions at every n tree #2175)
• Enable prediction on CPU-only machines with models trained with GPUs (multi-GPU training is not adaptable to other GPU counts or CPU #3342)
• Booster.load_model() should not require buffer writeability (Python: Booster.load_model() should not require buffer writeability #3013)
• Interoperability with sklearn.tree.export_graphviz (Interoperability with sklearn.tree.export_graphviz #2981)

JVM specific

• Update documentation
• Single-instance prediction interface (pull request for adding prdictInst API to xgboost4j #3153)
• Attach group info for each row in data frame ([jvm-packages] group data is only set for training set and is set incorrectly #3097)
• Evaluate an arbitrary number of evaluation sets when training in XGBoost-Spark ([jvm-packages] eval_set for xgboost4j-spark #3231)
• Pass additional parameters to XGBoostEstimator ([jvm-packages] xgboostj-spark: allow additional params to be set in XGBoostEstimator #3202)
• Do not expose booster parameter to XGBoost-Spark (gblinear issue in spark with objective as "reg":"linear","reg":"logistic" #3209)
• resolve external cache file for windows ([jvm-packages] Location of dtrain_cache files #3038)
• Float.NaN as missing value causes program crash ([jvm-packages] PySpark Support Checklist #3370)
• add early stopping in XGBClassifier/Regressor ([jvm-packages] Implemented early stopping #2710)
• Do train/test-split by column-values in XGBoost4J-Spark (Feature request: XGB-spark: Do train/test-split by column-values #3607)

Documentation

• Modernize XGBoost docs (Modernize XGBoost docs #3444)
• Document CSC/CSR format (Please improve DMatrix Java / CSR / CSC Documentation #3331)
• Document test system for developers (Roadmap: improve developer documentation #3283)
• Update the R tutorial "Discover your data" to reflect the deletion of RealCover (RealCover still present in documentation / vignettes. #3239)
• Document usage of C/C++ API

[hcho3]

Single instance prediction was added to JVM package by #3464.

[thvasilo]

Hello @hcho3 in the forum the ability to define the probability of each feature being selected when col_sample is being used has been requested a few times recently [1].

If we want to add this as a feature request I could create an issue to track it, and take it up since it doesn't seem too complicated once we agree on a design.

[1] https://discuss.xgboost.ai/t/sampsize-by-strata-in-subsample/281

[hcho3]

@thvasilo That would be great. Thanks!

[thvasilo]

Opened #3754 to track this.

[loretoparisi]

+1 for Multiple output regression