dedupe VS splink

I think this problem is known as data deduplication, in particular, entity deduplication. I googled a bit and it seems approaches vary from manual deduplication to some sort of active learning (if I am not mistaken). I am also curios if pre-trained transformer-based cross encoders can provide any good results (they are trained on sentences I think, but may be worth a try). Another problem here is how to measure progress (compare different approaches)?

I've had a good experience in the past with the dedupe package for these type activities. Unsure if it works for out-of-core type situations though, as my data set fit easily into memory.

Data deduplication is a super common problem, so it's useful experience to work on it. It's generally useful for companies, but I don't think it could be sold as a product unless is solving a very complicated, domain-specific de-duping problem. Otherwise, there are generic, open source de-duping tools such as: dedupe. It sounds like your model is similar to that.

Tools used: Python to scrape guest lists, dedupeio to better identify guests, Google Sheets to store and analyze the data, and Datawrapper to make the charts.

Try https://github.com/dedupeio/dedupe

https://github.com/moj-analytical-services/splink

Record linkage has been a big part of a project I've been working on for 6 months now. I personally think a great and free solution be using the splink package in Python which can handle 10+m rows which implements the Fellegi-Sunter model (equivalent to a naive-Bayes model) is the classical model in record linkage. It can be trained in an unsupervised manner using some initial parameter estimation (these are quite intuitive) and then expectation maximisation. The features in the model will be different pairwise string comparisons on your field of interest. These can include exact equality; edit distance comparisons like Levensthein distance and Jaro-Winkler; and phonetic comparisons like soundex and double metaphone. The splink pacakge will handle training the model and then all the graph theory at the end to connect all your links into clusters. All the details you'll need are in the links. https://www.robinlinacre.com/probabilistic\_linkage/ https://moj-analytical-services.github.io/splink/

https://moj-analytical-services.github.io/splink/ is a FOSS python package (but it runs against your db using SQL).

If you're curious, I've written a FOSS record linkage library that executes everything as SQL. It supports multiple SQL backends including DuckDB and Spark for scale, and runs faster than most competitors because it's able to leverage the speed of these backends: https://github.com/moj-analytical-services/splink

Splink - a python library for probabilistic record linkage (fuzzy matching/entity resolution).

Splink is dramatically faster and works on much larger datasets than other open source libraries. I'm particularly proud of the fact we support multiple execution backends (at the moment, DuckDb Spark Athena and Sqlite, but additional adaptors are relatively straightforward to write).

We've had >4 million pypi downloads and it's used in government, academia and the private sector, often replacing extremely expensive proprietary solutions.

https://github.com/moj-analytical-services/splink

More info in blog posts here:

Splink is a SQL tool that can do this https://github.com/moj-analytical-services/splink

Probabilistic record matching model such as a Fellegi-Sunter. Check out the splink package in Python.

Main docs here: https://moj-analytical-services.github.io/splink

It'd be great to see Splink add value in this area! Do give us a shout if you have any questions. The best place to post is on the Github discussions: https://github.com/moj-analytical-services/splink/discussions