Getting ready

In the previous section, we assigned the same weightage for each class; that is, the categorical cross entropy loss is the same if the magnitude of difference between actual and predicted is the same, irrespective of whether it is for the prediction of a default or not a default.

To understand the scenario further, let's consider the following example:

In the preceding scenario, the cross-entropy loss value is just the same, irrespective of the actual value of default.

However, we know that our objective is to capture as many actual defaulters as possible in the top 10% of predictions when ranked by probability.

Hence, let's go ahead and assign a higher weight of loss (a weight of 100) when the actual value of default is 1 and a lower weightage (a weight of 1) when the actual value of default is 0.

The previous scenario now changes as follows:

 

Now, if we notice the cross entropy loss, it is much higher when the predictions are wrong when the actual value of default is 1 compared to the predictions when the actual value of default is 0.

Now that we have understood the intuition of assigning weightages to classes, let's go ahead and assign weights to output classes in the credit default dataset.

All the steps performed to build the dataset and model remain the same as in the previous section, except for the model-fitting process.