Beginning Anomaly Detection Using

Figure 7-20.  The first five entries of the data frame

Figure 7-21.  The output in Figure 

7-20

 scrolled right

Chapter 7   temporal Convolutional networks

---

271

you can see that the data set is pretty massive with 284,807 entries in total (the index 

starts at 0). Additionally, notice how the values for time become absurdly large. If you 

pass in values this large into the model for training, you are bound to get errors with 

convergence. Not only that, it’s just good practice to normalize any large values, since it 

improves performance and training efficiency if you pass in smaller values to the model. 

Run the code in Figure 

7-23

 to standardize the values for Time and for Amount.

Now you can see that the values for the columns Time (Figure 

7-24

)

Figure 7-22.  The tail end of the data frame. Notice how large the values for 

time get

df['Amount'] = 

StandardScaler().fit_transform(df['Amount'].values.reshape(-1, 1))

df['Time'] = StandardScaler().fit_transform(df['Time'].values.reshape(-

1, 1))

df.tail()

Figure 7-23.  This code standardizes the values for Time and Amount

Chapter 7   temporal Convolutional networks

---

272

and for Amount (Figure 

7-25

)

are much smaller and much more manageable numbers to pass in.

Since there are so many entries in the entire data set, it’s best to limit the number 

of “normal” data entries you feed into the model since the model seems to ignore the 

anomalies if the entire data set is passed in. To avoid drowning out the anomalous data 

entries, let’s pick 10,000 normal entries to derive your training and testing data sets from 

(see Figure 

7-26

).

The output should look somewhat like Figure 

7-27

.