I'm doing a binary classification using K-nearest neighbor method. However when I look at the training data, 80% of the data is from category 2, and only 20% is from category 1.
Is it bad when you try to learn from that data? I feel like most of the prediction go to category 2.
Good or bad depends on your criterion.
Assuming your samples aren't biased, that is your out-of-sample has similar 80% C2 and 20% C1. (If not, then you should consider improve your data.) If your goal is to minimise the misclassification error, then predicting everything is category 2 isn't bad at all -- 80% success rate. This gives highest P1 = P(predict C1|true C1) = 1 and lowest P2 = P(predict C2|true C2) = 0
However, you can assign different weights to different categories if you want. So you do a weighted vote instead of a simple majority vote in you K neighbours for prediction. This would increase your performance on P2 with sacrifice on P1
