Beyond Machine Learning Pipelines with MLJ

This article introduces MLJ, an open-source machine learning toolbox developed in the Julia programming language by The Alan Turing Institute. Launched in late 2018, MLJ aims to unify the fragmented ecosystem of Julia machine learning packages by providing a consistent interface for both supervised and unsupervised models. The post highlights key features such as automatic hyperparameter tuning, homogeneous model ensembling, and a registry for model metadata that facilitates task-based model selection. A central focus is the introduction of 'learning networks,' a advanced composition feature that extends beyond traditional linear pipelines, allowing for complex model structures like stacking. The framework is designed to be data container agnostic, supporting Tables.jl formats, and includes a clean API for probabilistic modeling. While currently wrapping a limited number of native models, the project plans to integrate deep learning via Flux.jl and temporarily support scikit-learn models. The article serves as both a technical overview and a community invitation, encouraging developers to contribute to the repository, which has rapidly become the Institute's most-starred software project.