Beginning Anomaly Detection Using

ROC curve with AUC = 0.75 (Figure 

2-4

)

Figure 2-4.  ROC curve with AUC = 0.75

Chapter 2   traditional Methods of anoMaly deteCtion

---

32

The value for the AUC indicates that the model correctly predicts labels for data points 

only 75% of the time. It’s not bad, but it’s not good, so there’s clearly room to improve.

ROC curve with AUC = 0.5 (Figure 

2-5

)

The value for the AUC indicates that the model only has a 50% chance, or a 

probability of 0.5, to predict the correct label. This is about the worst AUC value  

you can get, since it means the model cannot distinguish between the positive and 

negative classes.

ROC curve with AUC = 0.25 (Figure 

2-6

)

Figure 2-5.  ROC curve with AUC = 0.5

Chapter 2   traditional Methods of anoMaly deteCtion

---

33

In this case, the model only has a probability of 0.25 to predict the right label, but this 

just means that it has a 0.75 probability of predicting the incorrect label. In the case that 

the AUC is 0, this means that the model is perfect at predicting the wrong label, meaning 

the labels are switched. If the AUC is < 0.5, this means the model gets better at predicting 

incorrectly as the AUC approaches 0.0. It’s the perfectly opposite case of when the AUC is 

> 0.5, where the model gets better at predicting correctly as the AUC approaches 1.0.

In any case, you want the AUC to be > 0.5, and at least greater than 0.9 and ideally 

greater than 0.95.

Figure 2-6.  ROC curve with AUC = 0.25

Chapter 2   traditional Methods of anoMaly deteCtion

---

34