robeespi/Fraud-detection-on-accounting-financial-systems ? reverse-engineered prompt

Build me a Google Colab friendly notebook that uses the included accounting fraud dataset to detect suspicious financial records.

I want it to walk through the whole flow in a way I can actually follow, starting with data prep and feature cleanup, then showing how fraud patterns look in the data with PCA so I can see the difference between global anomalies and local anomalies. After that, train a simple neural network model for the semi supervised case where only a small share of fraud examples are labeled, and also try a weakly supervised version where some fraud types are still unknown. If it makes sense, compare those results with an unsupervised autoencoder baseline.

Please keep it practical and readable, with clear notebook sections, comments, and basic explanations of what each step is doing. Show evaluation with F1 score and any useful plots, and add a t SNE style visualization if possible. Make sure it can run in Colab and note that it may need high RAM. Look up current docs online if you need to.