ML.NET 0.3 Release Notes

Today we are releasing ML.NET 0.3. This release focuses on adding components to ML.NET from the internal codebase (such as Factorization Machines, LightGBM, Ensembles, and LightLDA), enabling export to the ONNX model format, and bug fixes.

Installation

ML.NET supports Windows, MacOS, and Linux. See supported OS versions of .NET Core 2.0 for more details.

You can install ML.NET NuGet from the CLI using:

dotnet add package Microsoft.ML

From package manager:

Install-Package Microsoft.ML

Release Notes

Below are some of the highlights from this release.

• Added Field-Aware Factorization Machines (FFM) as a learner for binary classification (#383)

  • FFM is useful for various large sparse datasets, especially in areas such as recommendations and click prediction. It has been used to win various click prediction competitions such as the Criteo Display Advertising Challenge on Kaggle. You can learn more about the winning solution here.
  • FFM is a streaming learner so it does not require the entire dataset to fit in memory.
  • You can learn more about FFM here and some of the speedup approaches that are used in ML.NET here.

• Added LightGBM as a learner for binary classification, multiclass classification, and regression (#392)

  • LightGBM is a tree based gradient boosting machine. It is under the umbrella of the DMTK project at Microsoft.
  • The LightGBM repository shows various comparison experiments that show good accuracy and speed, so it is a great learner to try out. It has also been used in winning solutions in various ML challenges.
  • This addition wraps LightGBM and exposes it in ML.NET.
  • Note that LightGBM can also be used for ranking, but the ranking evaluator is not yet exposed in ML.NET.

• Added Ensemble learners for binary classification, multiclass classification, and regression (#379)

  • Ensemble learners enable using multiple learners in one model. As an example, the Ensemble learner could train both FastTree and AveragedPerceptron and average their predictions to get the final prediction.
  • Combining multiple models of similar statistical performance may lead to better performance than each model separately.

• Added LightLDA transform for topic modeling (#377)

  • LightLDA is an implementation of Latent Dirichlet Allocation which infers topical structure from text data.
  • The implementation of LightLDA in ML.NET is based on this paper. There is a distributed implementation of LightLDA here.

• Added One-Versus-All (OVA) learner for multiclass classification (#363)

  • OVA (sometimes known as One-Versus-Rest) is an approach to using binary classifiers in multiclass classification problems.
  • While some binary classification learners in ML.NET natively support multiclass classification (e.g. Logistic Regression), there are others that do not (e.g. Averaged Perceptron). OVA enables using the latter group for multiclass classification as well.

• Enabled export of ML.NET models to the ONNX format (#248)

  • ONNX is a common format for representing deep learning models (also supporting certain other types of models) which enables developers to move models between different ML toolkits.
  • ONNX models can be used in Windows ML which enables evaluating models on Windows 10 devices and taking advantage of capabilities like hardware acceleration.
  • Currently, only a subset of ML.NET components can be used in a model that is converted to ONNX.

Additional issues closed in this milestone can be found here.

Acknowledgements

Shoutout to pkulikov, veikkoeeva, ross-p-smith, jwood803, Nepomuceno, and the ML.NET team for their contributions as part of this release!