Categorisation

Categorisation accuracy is a system property, not a model property. I treat it as a craft: the design of the whole system that produces a result a business can use, not the tuning of one model.

The usual response is to improve a single model's score. In transaction data especially, that misses how the work actually happens. Accuracy comes from how different methods are combined, the order they run in, and how the data is shaped and reprocessed across more than one pass.

The systems I have designed treat categorisation as a pipeline of complementary methods: deterministic rules where the business already knows the answer, machine learning where the pattern is statistical, and later passes that use what earlier passes have already made reliable. The aim is not a more impressive model. It is a result the rest of the business can use.