Threshold adjustment
The decision threshold is a very important concept to keep track of. By default, we have the following:
- Prediction probability >= 0.5 implies Class 1
- Prediction probability < 0.5 implies Class 0
However, the threshold is a powerful meta-parameter that we are free to adjust. Table 5.3 shows predictions from a model versus the true labels.
If we use the default threshold of 0.5, the accuracy is 2/4 = 50%. If, on the other hand, the threshold chosen is 0.80, the accuracy is 100%. This shows how important the chosen threshold can be:
|
Predicted Output |
True Output |
|
0.65 |
0 |
|
0.75 |
0 |
|
0.85 |
1 |
|
0.95 |
1 |
Table 5.3 – A table showing the predicted...
The rest of the chapter is locked
