kmasiak/FraudDetection

The files in this repository were used for detecting accounting fraud using VAE-GAN and other models. Here is a breakdown of the files:

20220409-21_35_52_ep_3_decoder_model.pth - Decoder I trained that has the best results. 20220409-21_35_52_ep_3_discriminator_model.pth - Discriminator I trained that has the best results. 20220409-21_35_52_ep_3_encoder_model.pth - Encoder I trained that has the best results. Dataset.csv - The dataset used for train/testing, contains 9 features, 532909 regular, 70 global, and 30 local transactions. Fraud_Detection_AutoML.ipynb - AutoSklearnClassifier (an implementation of automl) is used on the fraud detection dataset. Fraud_Detection_Supervised.ipynb - KNN classifier is used on the fraud detection dataset. Gradio_Demo.ipynb - Note this is just for demo purposes. The actual implementation of the VAE-GAN model is not used in the gradio demo due to time constraints. SMOTE_VAE_GAN.ipynb - Use SMOTE to help mitigate the issue of an unbalanced dataset while training. VAE_GAN_Test.ipynb - Evaluates a VAE-GAN model. VAE_GAN_Train.ipynb - Trains the VAE-GAN model on the fraud detection dataset. ep_100_decoder_model.pth - Pre-trained decoder from a previous paper I used to improve results. ep_100_discriminator_model.pth - Pre-trained discriminator from a previous paper I used to improve results. ep_100_encoder_model.pth - Pre-trained encoder from a previous paper I used to improve results. Note: Credit for the above 3 files goes to Credit goes to Jie Dai, Chenjian Wang, and Shuoyi Wei. Accounting Fraud Detection with VAE-GAN, 2020.