The confusion matrix for a classification model compares the counts of the predicted class values with the observed or true class values.
It is a k x k table where k is the number of classes in the response. A good classifier has most of the counts on the diagonal from upper left to lower right (correct classifications) and few values on the off diagonal.