Photo of DeepakNess DeepakNess

Stripe's new transfer-based model

Unproofread notes

Stripe has developed a new approach to analyze transactions using a new transformer-based foundation model. Earlier, they relied on a traditional machine learning model but these models had limitations, but the new model is supposed to even increase the conversion even more and significantly decrease the fraudulent transactions.

Gautam Kedia, an AI/ML engineer at Stripe, explained this in a detailed X post. He mentions:

So we built a payments foundation model—a self-supervised network that learns dense, general-purpose vectors for every transaction, much like a language model embeds words. Trained on tens of billions of transactions, it distills each charge’s key signals into a single, versatile embedding.

This approach improved our detection rate for card-testing attacks on large users from 59% to 97% overnight.

While I did have a loose knowledge of what a transformer is, I looked up its definition again to understand it better in the context of payments:

A Transformer is a type of neural network architecture that has revolutionized natural language processing (NLP) and is now being applied to other domains, as seen in the Stripe example. Its key innovation is the attention mechanism.

The attention mechanism allows the model to weigh the importance of different parts of the input sequence when processing any single part.

Further, I asked Gemini to explain this entire thing to me in a simpler words and here's how it explained:

Think of it like reading a book. An older model might read word by word and only remember the last few words. A Transformer, with its attention mechanism, can look back at earlier parts of the book to understand the meaning of the current sentence in the broader context. In the payment world, this means understanding the significance of a transaction not just in isolation, but in the context of previous transactions.

Very cool.

Comment via email