The map to machine learning that never retrain-from-scratch

What it does

This is an awesome-list, not a library — a curated index of resources on online machine learning, where models learn incrementally from streaming data rather than in batch. It organizes courses, books, blog posts, software, and papers across the field: linear models, decision trees, drift detection, anomaly detection, clustering, and more. Think of it as a syllabus for anyone whose data refuses to sit still.

The interesting bit

The paper section doubles as a timeline of the field — from 1996’s BIRCH and 2000’s Hoeffding trees through Pegasos, Vowpal Wabbit-era ad-click prediction, up to 2022 online clustering benchmarks. You can watch two decades of “how do we learn from one example at a time” evolve down a single page. The software section is also honest about scope: it notes that Spark Streaming does mini-batching rather than true online learning, and that scikit-learn’s partial_fit is mostly mini-batch in disguise.

Key highlights

• Curated courses and books, including lecture notes on online convex optimization and a book on streaming ML with MOA.
• Software split into modelling (River, Vowpal Wabbit, MOA, Flink ML, PySAD) and deployment (KappaML, chantilly, a Django plugin for River models).
• Papers organized by topic: linear models, SVMs, decision trees, unsupervised learning, drift detection, anomaly detection, metric learning, and more.
• Blog posts include Chip Huyen’s real-time ML pieces and Max Halford’s write-ups on evaluating online models correctly.
• Maintained under the online-ml org, the same home as River — so the curation has a stake in the ecosystem.

Caveats

• It’s a link list, not tooling — no code, no benchmarks, nothing to run.
• Some linked software is explicitly flagged as unmaintained (e.g., LIBOL), and the list itself doesn’t claim to be exhaustive.

Verdict

Worth a bookmark if you’re entering streaming or online ML and want the canonical reading path. If you were hoping for a library or a comparison of tools, this is the signpost, not the destination.