99.4% Accurate but still completely useless?

An ML model proudly reports 99.4% accuracy but can still be utterly useless when dealing with imbalanced datasets. The model's dashboard appears green, stakeholders smile, and a report is prepared – but the crucial question looms: "How much fraud did the model actually catch?" The answer is zero. This illustrates the "accuracy trap."

Consider a simulated dataset of 20,000 card transactions, where only 0.6% are fraudulent. Introducing a simplistic baseline model, "lazy_model," that always predicts "not fraud," results in a 99.4% accuracy rate, yet the model fails to detect any fraud. Its metrics are astounding: precision, recall, and F1 score all at 0%. This model is correct most of the time but useful none of the time; it catches no fraud and probably still demands a promotion.

However, precision and recall offer a more revealing picture. Precision asks, "Of everything flagged as fraud, how much was actually fraud?" A low precision means the system frequently blocks genuine customers, falsely declaring them as fraud. Recall, on the other hand, asks, "Of all the fraud that really happened, how much did we catch?" Low recall indicates the fraudsters can walk away with the money while the model boasts its excellent accuracy.

When a real model is trained using balanced class weights, its results appear less impressive initially: accuracy drops to 99.4%, precision to 3.1%, recall to 77.8%, and F1 score to 6.0%. Although the accuracy is lower, the real model successfully catches nearly 78% of the fraud. Thus, the question remains: which model is better?

The answer lies in the trade-off between precision and recall. Lowering the fraud threshold increases recall but raises false alarms, while raising the threshold improves precision but allows more fraud to slip through. The "best" threshold is a business decision depending on the cost of different errors. Is it more detrimental to annoy a legitimate customer or miss a fraudulent transaction? The answer varies based on the system.

In conclusion, for imbalanced classification problems, accuracy is not useless but often incomplete. It is essential to examine precision, recall, F1 score, and confusion matrix to understand the types of mistakes made. Lastly, consider the business impact of those mistakes. A 99.4% accurate model can still be terrible, and metrics can easily mislead us if not used carefully.