Max Halford

Migrate all mini-batch (_many) methods to narwhals for dataframe-agnostic support

2

## Motivation [#1900](https://github.com/online-ml/river/pull/1900) introduced dataframe-agnostic mini-batching via [narwhals](https://github.com/narwhals-dev/narwhals): inputs are wrapped at the method boundary, the numpy compute core stays untouched, and outputs are rebuilt in the caller's native backend...

FSRS

1

FSRS is the current best in class model for spaced-repetition learning. It's an online algorithm, with only a few parameters. The implementation that the Anki project uses is [here](https://github.com/open-spaced-repetition/fsrs-optimizer). There...

Rewrite LocalOutlierFactor as a windowed, on-demand detector

1

## What Rewrites `anomaly.LocalOutlierFactor` from the incremental-LOF bookkeeping (nine threaded dicts + a stack of free functions) into a small class that delegates storage and neighbor search to a `river.neighbors`...

# Explore: move `linear_model` mini-batch learning into Rust (true per-sample SGD) ## Background: `learn_many` is not equivalent to looping `learn_one` Today `linear_model`'s `GLM.learn_many` (used by `LinearRegression` / `LogisticRegression`) does **mini-batch...

Add online EWA and shrinkage covariance/precision estimators

1

## What Adds a family of **online covariance estimators** to `river.covariance`, reimplemented from the [`precise`](https://github.com/microprediction/precise) package following River conventions (dict-native, Narwhals mini-batches), per discussion #1884. Only genuinely-online methods are included...

Implement Bayesian logistic regression

4

## Motivation `linear_model.BayesianLinearRegression` maintains a full Gaussian posterior over the weights, which gives calibrated predictive uncertainty (`predict_one(..., with_dist=True)`) and makes it a natural fit for downstream uses like `bandit.LinUCB`. There...

## Summary Proposal to make `river.drift` a focused area of development for the next ~6 months: broaden detector coverage across the three main families (online statistical CPD, supervised/error-stream, and unsupervised/data...

Make AMF estimators drift resilient

4

AMFClassifier and AMFRegressor are the most accurate models we have. But they are not designed to do well with drift. Indeed, I realized when working on https://github.com/online-ml/river/pull/1825 that copying the...

It would be very cool to ship a minimal version of Python that can run on [MicroPython](https://micropython.org/). Personally I'd love to see that happen. There's also [CircuitPython](https://circuitpython.org/).

https://www.ijcai.org/Proceedings/11/Papers/364.pdf