Im using Keras on a multi-class classification problem. The categories in the data set are pretty qualitative and cause the classifier a lot of confusion. The predict_classes() function assigns classes to samples that correspond to a low (sometimes 60%) score from the predict() function.
With this particular problem, the absolute accuracy is less important than the number of false predictions. Meaning that when a category is predicted, I want it to be very certain that it is accurate. I dont care if 9/10 samples get dumped into an "uncertain" class as long as the 1/10 has a high accuracy.
Is there a way to create this "uncertain" class within the .predict_classes() function? Is there another way to achieve my goal? Is there a way to incorporate the "uncertain" class into the training process to decrease the number of false positives?
