Parameter tuning in classification

Question

Recently, I am digging into the selection of tuning parameters in a binary classification.

I gathered information by googling and the following is my organization.

1. We can distinct binary classification problems into two cases

• Balanced classification: the proportion of the two classes is about 5:5
• Imbalanced classification: the proportion of the two classes are not balanced

2. In imbalanced classification problems, "accuracy" is not a good metric for model performance

• Almost classification algorithms are developed considering the balanced classification, thus, if our data set is imbalanced the algorithm classifies almost observations as the dominate class.
• Therefore, although we cannot correctly classify the fewer proportion class, the accuracy is very high

3. Hence, when we tune our tuning parameters by using cross-validation, the accuracy is not a good metric

My questions are as follows:

1. Is there any wrong part in my knowledge?

2. If my thought is not critically wrong, I think, we can use the accuracy to tune the parameters in the "balanced" classification problems. But, I am not sure whether that is correct idea

3. In every classification, we have to choose a probability threshold (= decision threshold (cutoff) = discriminant threshold (cutoff)). Considering the discussion above, it is not a good idea to use "accuracy" to select the threshold as a metric only in the imbalanced classification problems. I think, however, accuracy could be a good metric in the threshold selection in a balanced classification problem. Is this idea non-problematic?

Thank you for your time to read this long question.