Beginning Anomaly Detection Using

Figure 4-21b.  Code to show the precision and recall for threshold = 1.0

Figure 4-21c.  Code to show the precision and recall for threshold = 5.0

Figure 4-21d.  Code to show the precision and recall for threshold = 15.0

If you observe the four classification reports, you can see that the precision and recall 

columns are not good (note the very low values for precision and recall in row 0 and row 1)  

for threshold = 1 or 5. They look better for threshold = 10 or 15. In fact, threshold = 10 

looks pretty good with a good recall and also higher precision than for threshold = 1 or 5.

Threshold = 5.0

Threshold = 15.0

Chapter 4   autoenCoders

---

139

Picking a threshold is a matter of experimentation in this and other models and 

changes as per the data being trained on.

Compute the AUC (Area Under the Curve, 0.0 to 1.0) which comes up as 0.86. 

Figure 

4-21e

 shows the code to show AUC.

Figure 4-21e.  Code to show AUC

You can now visualize the confusion matrix to see how well you did with the model. 

Figure 

4-22

 shows the confusion matrix.

Figure 4-22.  Confusion matrix

Now, using the predictions of the labels (normal or anomaly), you can plot the 

anomalies in comparison to the normal data points. Figure 

4-23

 shows the anomalies 

based on the threshold.

Chapter 4   autoenCoders

---

140