SKLearn的演进之路

本文档的目的

This document list general directions that core contributors are interested to see developed in scikit-learn. The fact that an item is listed here is in no way a promise that it will happen, as resources are limited. Rather, it is an indication that help is welcomed on this topic.

Statement of purpose: Scikit-learn in 2018

Eleven years after the inception of Scikit-learn, much has changed in the world of machine learning. Key changes include:

  • Computational tools: The exploitation of GPUs, distributed programming frameworks like Scala/Spark, etc.
  • High-level Python libraries for experimentation, processing and data management: Jupyter notebook, Cython, Pandas, Dask, Numba…
  • Changes in the focus of machine learning research: artificial intelligence applications (where input structure is key) with deep learning, representation learning, reinforcement learning, domain transfer, etc.

A more subtle change over the last decade is that, due to changing interests in ML, PhD students in machine learning are more likely to contribute to PyTorch, Dask, etc. than to Scikit-learn, so our contributor pool is very different to a decade ago.

Scikit-learn remains very popular in practice for trying out canonical machine learning techniques, particularly for applications in experimental science and in data science. A lot of what we provide is now very mature. But it can be costly to maintain, and we cannot therefore include arbitrary new implementations. Yet Scikit-learn is also essential in defining an API framework for the development of interoperable machine learning components external to the core library.

Thus our main goals in this era are to:

  • continue maintaining a high-quality, well-documented collection of canonical tools for data processing and machine learning within the current scope (i.e. rectangular data largely invariant to column and row order; predicting targets with simple structure)
  • improve the ease for users to develop and publish external components
  • improve inter-operability with modern data science tools (e.g. Pandas, Dask) and infrastructures (e.g. distributed processing)

Many of the more fine-grained goals can be found under the API tag on the issue tracker.

Architectural / general goals

The list is numbered not as an indication of the order of priority, but to make referring to specific points easier. Please add new entries only at the bottom.

  1. Everything in Scikit-learn should conform to our API contract
    • Pipeline <pipeline.Pipeline> and FeatureUnion modify their input parameters in fit. Fixing this requires making sure we have a good grasp of their use cases to make sure all current functionality is maintained. :issue:`8157` :issue:`7382`
  2. Improved handling of Pandas DataFrames and SparseDataFrames
    • document current handling
    • column reordering issue :issue:`7242`
    • avoiding unnecessary conversion to ndarray
    • returning DataFrames from transformers :issue:`5523`
    • getting DataFrames from dataset loaders
    • Sparse currently not considered
  3. Improved handling of categorical features
    • Tree-based models should be able to handle both continuous and categorical features :issue:`4899`
    • In dataset loaders
    • As generic transformers to be used with ColumnTransforms (e.g. ordinal encoding supervised by correlation with target variable)
  4. Improved handling of missing data
    • Making sure meta-estimators are lenient towards missing data
    • Non-trivial imputers
    • Learners directly handling missing data
    • An amputation sample generator to make parts of a dataset go missing
    • Handling mixtures of categorical and continuous variables
  5. Passing around information that is not (X, y): Sample properties
    • We need to be able to pass sample weights to scorers in cross validation.
    • We should have standard/generalised ways of passing sample-wise properties around in meta-estimators. :issue:`4497` :issue:`7646`
  6. Passing around information that is not (X, y): Feature properties
    • Feature names or descriptions should ideally be available to fit for, e.g. . :issue:`6425` :issue:`6424`
    • Per-feature handling (e.g. “is this a nominal / ordinal / English language text?”) should also not need to be provided to estimator constructors, ideally, but should be available as metadata alongside X. :issue:`8480`
  7. Passing around information that is not (X, y): Target information
    • We have problems getting the full set of classes to all components when the data is split/sampled. :issue:`6231` :issue:`8100`
    • We have no way to handle a mixture of categorical and continuous targets.
  8. Make it easier for external users to write Scikit-learn-compatible components
    • More flexible estimator checks that do not select by estimator name :issue:`6599` :issue:`6715`
    • Example of how to develop a meta-estimator
    • More self-sufficient running of scikit-learn-contrib or a similar resource
  9. Support resampling and sample reduction
  10. Better interfaces for interactive development
  11. Improved tools for model diagnostics and basic inference
    • partial dependence plots :issue:`5653`
    • alternative feature importances implementations (e.g. methods or wrappers)
    • better ways to handle validation sets when fitting
    • better ways to find thresholds / create decision rules :issue:`8614`
  12. Better tools for selecting hyperparameters with transductive estimators
    • Grid search and cross validation are not applicable to most clustering tasks. Stability-based selection is more relevant.
  13. Improved tracking of fitting
    • Verbose is not very friendly and should use a standard logging library :issue:`6929`
    • Callbacks or a similar system would facilitate logging and early stopping
  14. Use scipy BLAS Cython bindings
    • This will make it possible to get rid of our partial copy of suboptimal Atlas C-routines. :issue:`11638`
    • This should speed up the Windows and Linux wheels
  15. Allow fine-grained parallelism in cython
    • Now that we do not use fork-based multiprocessing in joblib anymore it’s possible to use the prange / openmp thread management which makes it possible to have very efficient thread-based parallelism at the Cython level. Example with K-Means: :issue:`11950`
  16. Distributed parallelism
    • Joblib can now plug onto several backends, some of them can distribute the computation across computers
    • However, we want to stay high level in scikit-learn
  17. A way forward for more out of core
    • Dask enables easy out-of-core computation. While the dask model probably cannot be adaptable to all machine-learning algorithms, most machine learning is on smaller data than ETL, hence we can maybe adapt to very large scale while supporting only a fraction of the patterns.
  18. Better support for manual and automatic pipeline building
  19. Support for working with pre-trained models
    • Estimator “freezing”. In particular, right now it’s impossible to clone a CalibratedClassifierCV with prefit. :issue:`8370`. :issue:`6451`
  20. Backwards-compatible de/serialization of some estimators
    • Currently serialization (with pickle) breaks across versions. While we may not be able to get around other limitations of pickle re security etc, it would be great to offer cross-version safety from version 1.0. Note: Gael and Olivier think that this can cause heavy maintenance burden and we should manage the trade-offs. A possible alternative is presented in the following point.
  21. Documentation and tooling for model lifecycle management
    • Document good practices for model deployments and lifecycle: before deploying a model: snapshot the code versions (numpy, scipy, scikit-learn, custom code repo), the training script and an alias on how to retrieve historical training data + snapshot a copy of a small validation set + snapshot of the predictions (predicted probabilities for classifiers) on that validation set.
    • Document and tools to make it easy to manage upgrade of scikit-learn versions:
      • Try to load the old pickle, if it works, use the validation set prediction snapshot to detect that the serialized model still behave the same;
      • If joblib.load / pickle.load not work, use the versioned control training script + historical training set to retrain the model and use the validation set prediction snapshot to assert that it is possible to recover the previous predictive performance: if this is not the case there is probably a bug in scikit-learn that needs to be reported.
  22. (Optional) Improve scikit-learn common tests suite to make sure that (at least for frequently used) models have stable predictions across-versions (to be discussed);
    • Extend documentation to mention how to deploy models in Python-free environments for instance ONNX. and use the above best practices to assess predictive consistency between scikit-learn and ONNX prediction functions on validation set.
    • Document good practices to detect temporal distribution drift for deployed model and good practices for re-training on fresh data without causing catastrophic predictive performance regressions.
  23. More didactic documentation
    • More and more options have been added to scikit-learn. As a result, the documentation is crowded which makes it hard for beginners to get the big picture. Some work could be done in prioritizing the information.

Subpackage-specific goals

sklearn.cluster

  • kmeans variants for non-Euclidean distances, if we can show these have benefits beyond hierarchical clustering.

sklearn.ensemble

  • a stacking implementation
  • a binned feature histogram based and thread parallel implementation of decision trees to compete with the performance of state of the art gradient boosting like LightGBM.

sklearn.model_selection

  • multi-metric scoring is slow :issue:`9326`
  • perhaps we want to be able to get back more than multiple metrics
  • the handling of random states in CV splitters is a poor design and contradicts the validation of similar parameters in estimators.
  • exploit warm-starting and path algorithms so the benefits of EstimatorCV objects can be accessed via GridSearchCV and used in Pipelines. :issue:`1626`
  • Cross-validation should be able to be replaced by OOB estimates whenever a cross-validation iterator is used.
  • Redundant computations in pipelines should be avoided (related to point above) cf daskml

sklearn.neighbors

  • Ability to substitute a custom/approximate/precomputed nearest neighbors implementation for ours in all/most contexts that nearest neighbors are used for learning. :issue:`10463`

sklearn.pipeline

  • Performance issues with Pipeline.memory
  • see “Everything in Scikit-learn should conform to our API contract” above
  • Add a verbose option :issue:`10435`