Transaction categorization can seem like a tradeoff between two subpar choices. On the one hand, you can attempt to build an in-house machine learning solution, which requires time, money, and effort - only to produce something that is likely still brittle and non-performant. On the other hand, you can opt for external solutions. What you gain in time though, you lose in control.
These respective solutions are incomplete. Expecting that your specific use case matches the general outputs of an external provider, with no loss in translation, is a tall-order. But, what if you could have it both ways? What if you could combine the power of an external solution like Ntropy Core Model (trained on millions of examples of diverse transactions), with the deftness of specialized models, built for your particular use cases?
Every Ntropy Custom Model is built using the Ntropy Core Model, and can reach 90% accuracy with as few as six (6) samples per label. Our Core Model, which we released in Q2 of 2021, is the leading API solution for general transaction categorization. It utilizes collective knowledge from other transactions in the network, has access to merchant databases and search engines, understands natural language, and uses the context of a transaction’s history. Building Custom Models on top of our Core Model allows everyone to benefit from the breadth of transaction types that the Ntropy Core Model sees, which any specialized in-house model would never have access to.
By opening up our Core categorization model to specialized in-house models, the power of a Custom Model becomes 1+1 = 3.
In this post, we explain the key improvements from customization, and how to easily get started:
One hundred percent (100%) general categorization accuracy is impossible. No matter how much you pay your ML team, you’ll never be able to train a model to think an “intra-account transfer” includes “wire transfers,” but then to predict that “wire transfers” are not “intra-account transfers.” With a custom approach, we allow you to train hyper-specific models that understand what each individual category means to your use case. Translation: by using an Ntropy Custom Model, you can expect to see at least a 19% improvement in accuracy over performance from any general categorization API.
There is no unique way to analyze transaction data, so why use one model for every use case? The transaction data you require to build one product, can be used for multiple products, features or use cases. With Ntropy Custom Models you have the flexibility to run more than one model at a time, providing diverse views of the same data set, and allowing for deeper insight into what transactions mean. One model can predict that restaurant purchases are an “employee expense” while the second model can predict that they are a “customer acquisition cost.” As your categorization needs evolve, so too should your model.
Performance shouldn’t have to depend on Ntropy. When data changes, you should be able to change with it. Customization allows for users to quickly make their own updates to a model without needing to wait for the Ntropy Core Model to update. In turn, there’s no need to make updates only when data changes. Ntropy Custom Models let you set the performance levels you want and achieve them by controlling how and when you update your model.
Last but not least, we believe that the best products are simple products. With Ntropy Custom Models, we have condensed the training, testing, and deployment into only 2-3 lines of code. And because customization is built on top of our Core Model, any updates will automatically improve your Custom Models - without any additional effort 🏖.
With improved accuracy, flexibility, and speed to deployment, our goal is to make customization the new norm. Whether you know exactly what you need or are still exploring, we have a library of resources below to help you get there. We can’t wait to see what you build!
To get started, request an API key and more information CONTACT US HERE. We’ll get back to you faster than we can train your custom model.
If you have labeled data, check our Quickstart Colab here for a deeper look into deployment.
If you don’t have labeled data, or aren’t sure what to look for yet, we’ve got you covered; you can still get started with our Core Model. To see how, check out our Tutorial Colab
If you still want to learn more about customization this blog post will tell you everything you need to know.