Discussion questions#
The model outputs: Happy = 0.22; Sad = 0.15l Angry = 0.55; Afraid = 0.08. Is this person “angry”? Did the model discover an emotion or impose one? What is lost when we force a continuous psychological state into discrete categories? In what ways could multi-class (>2) classification tell us more than a label?
Imagine a system that predicts whether or not someone has a disease that is expensive to treat. Lowering the threshold catches more real cases, but also falsely flags many patients. Raising the threshold avoids false alarms, but misses some real at-risk people. What factors should we consider when informing what threshold would be best in this scenario? Are there certain sources of evidence that we should value more when considering our decision?