Dataset creation
Closed, ResolvedPublic
Actions

Assigned To

Authored By

	• XiaoXiao-WMF
	Jan 6 2025, 3:01 PM

Project Tags

Referenced Files

None

Subscribers

Description

This task consists of three steps:

Based on the data needed for the identifier, we would like to create a reproducible approach for creating a dataset that takes all available information (some db only goes back 3 months) per edit (list not final):
- All collected client hints
- Revision info, e.g. revision id, timestamp, actor name, ip, etc.
- IPoid info: https://office.wikimedia.org/wiki/User:KHarlan_(WMF)/OpenSearch_ipoid

Create weak labels - label reported sock puppets from this page: https://en.wikipedia.org/wiki/Wikipedia:Sockpuppet_investigations#Cases_currently_listed_at_SPI and incorporate to the dataset created in step 1
- Having weak labeling will greatly inform weight params of the LSH algo, as well as for testing
- Note: Confirmed cases are here https://en.wikipedia.org/wiki/Category:Wikipedia_sockpuppets, but most are not within the 3-month time frame, please double check

(Experimental) Create ban-invasion dataset use (potentially) graph algo
- If successful, it will be used as an additional testing dataset

		Status	Subtype	Assigned	Task
		Resolved		• XiaoXiao-WMF	T381031 WE4.2.10 Add more browser signals to client hints pipeline to generate unique device identifier
		Resolved		fkaelin	T383060 Dataset creation

• XiaoXiao-WMF set Due Date to Feb 7 2025, 5:00 AM.

• XiaoXiao-WMF changed the task status from Open to In Progress.Jan 6 2025, 5:04 PM

• XiaoXiao-WMF triaged this task as High priority.

• XiaoXiao-WMF updated the task description. (Show Details)

Weekly updates:

Focused on research iteration, by building blocks in a notebook for experimentation that can run end-to-end
Pipeline to extract client hints table from mariadb tables (all available cu_changes data for the 20 largest wikis). In prod this would be scheduled as airflow dag .
Enrich with ip based features (city/country/continent & isp) using maxmind
Compute the fingerprint hashes based on https://github.com/xiao-xiao-h/browser-fingerprint-lsh/tree/main
Implementations for other feature encoding algorithms: one-hot encoding, generic feature hashing, and label embeddings
Support for dimensionality reduction using PCA/SVD
Support for nearest neighbor search using a faiss index, using IndexLSH
Methods to train indeces (with optional dimensionality reduction) using different encoding methods on a training dataset
Evaluation helpers to run inference and display results on the test dataset
Datasets in hdfs folder /user/fab/we421. included wiki_db: arwiki, arzwiki, cawiki, dewiki, enwiki, eswiki, fawiki, frwiki, idwiki, itwiki, jawiki, nlwiki, plwiki, ptwiki, ruwiki, svwiki, ukwiki, viwiki, warwiki, zhwiki
Code is available here