The Fifth Elephant
The Fifth Elephant

The Fifth Elephant 2026 Winter Edition

Call for Data Problems and Proposals

All submissions

Reginthala Mahesh

@Mahesh1735

Who Else Is This Customer? Linking Fraud Accounts Without Labels, in Pictures

Submitted Oct 9, 2026

🎤 I am submitting: To speak I have a submission for: Poster - work I have done that I can share in the alleyways with attendees

Submission type: Poster - alleyway

The poster

When a fraudster’s account gets blocked, they simply make a new one. This poster starts from one real (anonymised) fraud ring of 611 accounts and shows, mostly in pictures, how we found it. We took about 27 million accounts, plus the devices, phone numbers, UPI IDs, locations and names they share, and grouped together the accounts that belong to the same real person. We did this without a single labelled example of “these two accounts are the same person”.

The poster has four panels. Panel 1 shows the data as a map, with accounts on one side and shared things on the other, plus the cleaning traps we hit, such as names typed in decorative Unicode letters and “ground floor” used as a building name. Panel 2 shows how much each kind of shared signal really proves, from a device (strong) to an IP address (weak), learned from the data with a record-linkage method from 1969. Panel 3 puts three methods side by side and shows the surprise: a cheap counting method matched an expensive graph-embedding model. Panel 4 shows how we judged the results with no labels, and what replaying past orders says about savings: about 20 paise per order on one day, and 12 paise per order on another.

Come and vote

The poster will include six anonymised borderline clusters. Visitors can vote: is this one person, or many different people? Those votes become a small labelled set for the problem we find hardest, which is telling a fraud ring apart from a college hostel.

Takeaways

  1. Not all sharing is equal. Sharing a rare device is strong proof that two accounts are one person. Sharing a common IP address proves almost nothing. Weighting evidence by rarity took us as far as a much more expensive model.
  2. No labels? Decide what “good” should look like, then try to break it. Write down the pattern real results should show, test it on data the model never saw, and ask how your check could be rewarding the wrong thing.

Who will benefit most

Fraud, risk and trust-and-safety teams, anyone deduplicating customer or patient records, KYC and lending teams, and anyone doing data science without labels. No graph or ML background needed.

About the data

All identifiers on the poster are masked. Aggregate numbers are real. The raw data stays private because it contains personal information.

Bio

Mahesh Reginthala is a Senior Machine Learning Engineer at Zepto in Bengaluru, He started and led the identity resolution work on this poster, taking it from an idea to a weekly pipeline that groups about 27 million accounts across 67 cities. He also started Zap, Zepto’s in-house AI support system, and earlier built AI ad-creative generation at Walmart and search ranking at Udaan. He studied at IIT Guwahati. The identity resolution system was built together with Prema Chandra Reddy and the Fraud Data Science team at Zepto.

Draft poster

Not ready yet. I will share a draft layout, with comments on, before the editorial call.

What I don’t know yet

  • How to check identity clusters without ground truth. Our main check uses profit and refunds per order. A method can look good on that check by grouping costly customers who are not the same person. I would love ideas for better checks.
  • Shared living versus shared identity. Hostels, offices and joint families create big clusters of genuinely different people. How do others draw that line?

Match-making tags

  • I can help with: experience, technique, critique
  • I need help with: technique, critique, dataset
  • I’d like to meet: practitioners, researchers, domain experts, dataset owners

Topic tags

fraud, entity resolution, evaluation

All submissions

Comments

{{ gettext('Login to leave a comment') }}

{{ gettext('Post a comment…') }}
{{ gettext('New comment') }}
{{ formTitle }}

{{ errorMsg }}

{{ gettext('No comments posted yet') }}

Hosted by

The Fifth Elephant
The Fifth Elephant
Jumpstart better data engineering and AI futures