Transaction classifier

Type a retail transaction description. The page runs the trained TF-IDF, SVD and logistic regression model in JavaScript and returns a category with its top-3 probabilities. Nothing is sent to a server.

How good is it?

The dataset has no category labels. The categories are 56 KMeans clusters, named after their heaviest terms, and the classifier learns to reproduce them. Its accuracy against those cluster pseudo-labels measures consistency with the clustering, not real-world correctness. On 50 descriptions labelled by hand, the top-1 category was right 62% of the time when a fitting category existed (85% top-3), and 16 of the 50 had no fitting category at all. Details are in the README.