Beginning Anomaly Detection Using

training data set (the data used to train the model so that it can learn to classify 

anomalies and normal data). This is because in the real world, there is the factor of 

unpredictability that even has humans confused at times. The world would be a simpler 

place if data is black and white, so to speak, but more often than not, there is a huge 

Figure 2-3.  ROC curve with AUC = 1.0

Chapter 2   traditional Methods of anoMaly deteCtion

---

31

gray area (are we sure that point is X and not Y? Is this really an anomaly or just a really 

weird case of a normal point?). For deep learning models, it is important that they keep 

achieving high AUC scores when exposed to new data that includes plenty of variation. 

Basically, it’s a reasonable assumption to expect a slight drop in performance when 

exposing your model to new data outside of your training set.

The goal with training models is to avoid overfitting and to keep the AUC as high as 

possible. If the AUC turns out to be 0.99999 even after being exposed to an extremely 

large sample of new data that includes a lot of variety, that means the model is basically 

about as ideal of a model we can get and has far surpassed human performance, which 

is impossible for the time being.