Which machine learning approach matches this data and goal?

Supervised learning trains on labeled data where the correct output is already known, and 400,000 investigated transactions are exactly that. The semi-supervised reading is the trap: incoming transactions do arrive without labels, but those are inputs the model scores, not training data, and semi-supervised learning describes a training set that is only partly labeled. The unsupervised reading is true about fraud - it is rare and it moves - and choosing it means discarding the strongest signal the company already owns. Reinforcement learning would fit if the model improved from a reward loop as analysts confirmed or rejected its flags; that is a plausible later design, not what a two-year labeled history calls for.