Given the class imbalance ratio, we recommend measuring the accuracy using the Area Under the Precision-Recall Curve (AUPRC). Confusion matrix accuracy is not meaningful for imbalanced classification. Regarding this, use linear machine learning models, such as random forests, logistic regression, or support vector machines, by applying over-or under-sampling techniques. Alternatively, we can try to find anomalies in the data, since an assumption like only a few fraud cases being anomalies within the whole dataset.

When dealing with such a severe imbalance of response labels, we also need to be careful when measuring model performance. Because there are only a handful of fraudulent instances, a model that predicts everything as non-fraud will already achieve more than the accuracy of 99%. But despite its high accuracy, linear machine learning models won't necessarily help us find fraudulent cases.

Therefore, it would be worth exploring deep learning models, such as autoencoders. Additionally, we need to use anomaly detection for finding anomalies. In particular, we will see how to use autoencoders to pre-train a classification model and measure model performance on unbalanced data.