online-ml/river

Online GapEncoder

Open

#1,439 opened on Nov 3, 2023

View on GitHub
 (1 comment) (0 reactions) (0 assignees)Python (4,574 stars) (553 forks)batch import
Good first issueNew feature

Description

skrub is a wonderful new project related to scikit-learn. You can see Gaël Varoquaux present it here. They have a transformer called GapEncoder: it's a way to embed fuzzy strings. This could be really powerful online, say for classifying Tweets or Twitch messages, where typos are aplenty.

We already have a way to do online TD-IDF/count vectorization. But we don't have Gamma-Poisson matrix factorization. It is doable online though. Once we have it, we could assemble the two into a nice GapEncoder class. See paper here.

This is related to #1412. Indeed, maybe this works well without Gamma-Poisson matrix factorization. For instance, we could use decomposition.LDA, which we already have.

Contributor guide

Tech stack
python
Domain
machine learningdata
Issue type
feature
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
5
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
over 1 week
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
needs maintainer response
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
mostly clear
Prerequisites
Gamma Poisson matrix factorizationonline learningPython
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
15
Research direction
The issue proposes implementing an online version of GapEncoder, which involves Gamma Poisson matrix factorization. Start by reviewing the referenced paper (https://inria.hal.science/hal 02171256v4) and the existing River implementation for online TF IDF. Consider adapting decomposition.LDA as suggested, or implement the Gamma Poisson model from scratch. Check related issue #1412 for potential overlap. Key files to look at are the decomposition module in River and the skrub GapEncoder source for reference. The main challenge is ensuring the factor update rules are online and computationally efficient.