Beginning Anomaly Detection Using

Figure 5-54.  The output shapes and some entries of y_test are displayed

Figure 5-55.  Initializing the model

The code to train the model is shown in Figure 

5-56

.

Figure 5-56.  Training the model on x_train, using x_test as validation data

Chapter 5   Boltzmann maChines

---

207

The output you should see is shown in Figure 

5-57

.

Figure 5-57.  The training output by the model for the code in Figure 

5-56

Since the labels aren’t binary, you want to redefine them as either normal, 0, or 

anomalous, 1. Run the code in Figure 

5-58

.

Chapter 5   Boltzmann maChines

---

208

The output you should see is shown in Figure 

5-59

.

Figure 5-58.  Code to change all labels that are 4 to 0, representing normal entries, 

and all labels that aren’t 4 to 1, representing anomalies

Figure 5-59.  The labels should now be transformed. Some of the entries in y_test 

are shown to make sure they were transformed correctly

Now that your labels have been corrected, you can get the free energy and find the 

AUC score (see Figure 

5-60

).

Figure 5-60.  Code to get the free energy for each model in x_test and then to find 

the AUC score based on that

Chapter 5   Boltzmann maChines

---

209

The output you should see is shown in Figure 

5-61

.

Figure 5-61.  The generated AUC score

That’s an even better AUC score than for the credit card data set! Let’s take a look 

at what happens when you plot the free energy vs. the probability. As with the previous 

example, let’s take a look at the five-number summary for the normal data to see how the 

distribution looks (Figure 

5-62

 and Figure 

5-63

).

Figure 5-62.  Code to check the five-number summary of the normal data

The output should look somewhat like Figure 

5-63

.

Figure 5-63.  It seems that the graph is skewed right, and that all of the values are 

under 1150

Chapter 5   Boltzmann maChines

---

210

Now let’s look at the five-number summary to see what the general distribution of 

the anomalous data looks like (see Figure 

5-64

 and Figure 

5-65

).