Steps followed in Notebook (for financial fraud detection in trasactions):
- First we understood the data from dataset, we have transactions details with flags whether they are fraud or not, we drawn pivot table for overall numbers in dataset
- We computed the difference between original and destination balances after a transaction, creating binary indicators based on whether the debited amount matches the change in balance.
- We tokenized customer names using TensorFlow's Tokenizer class and then padded the sequences to a fixed length of 1. The purpose of this process seems to be converting textual customer names into numerical representations for model input.
- We applied Logistic Regression, Decision Tree, k-Nearest Neighbors, Support Vector Classifier (SVC), Naive Bayes (GaussianNB), and Random Forest ML algorithms and evaluated them using a cross validation method.
- We showed a visual representation of the model's performance, breaking down the counts of true positives, true negatives, false positives, and false negatives, with an option for normalization.
- We chose best model which has low false positive counts and winner were Naive bayes and random forest.