Finance / Merchant payments
Square: Classifying the business category of merchants
- Company
- Square
- Country
- United States
- Adoption stage
- In operation
- Source published
- Date basis
- The date the source was published. It can differ from the date adoption started.
- How the source was checked
- Checked a source excerpt through a search tool
The work problem
Square mostly relied on the business category sellers chose for themselves when they signed up. That method was easy to get wrong. Some sellers picked carelessly to finish signup quickly, and some, such as a shop that does both hair and nails, found it hard to pick just one. The category code also affects the card fees Square pays and the benefits it receives from card networks.
Technology and data
Square built a category classification model on the RoBERTa architecture, using roberta-large. The training material is a random sample of more than 20,000 sellers reviewed by people. The input is the business name, the category and subcategory chosen at signup, and the list of items and services sold. Items are sorted by how often they are bought and cut down to a length the model can read. Items generated automatically at signup are dropped, leaving only entries the seller edited. Training ran on a GPU cluster at Databricks, and predictions are refreshed daily.
Results
Square estimated that the accuracy of category identification rose by about 30 percent in absolute terms over the previous method. In the test data the gain over self-selection was largest in retail and home repair. The company said it uses these predictions across every internal metric that needs a split by business category.
Limits and open questions
The accuracy improvement is an estimate from Square's own held-out test data. The source gives no figure showing that it led to lower fees or a change in payment volume.
Sources
- RoBERTa Model for Merchant Categorization at Squaredeveloper.squareup.com, Accessed
Compiled from public sources. These are not results from ATF Works customers.