Based on my understanding, we train a simple model (eg. Logistic Regression) to predict the engagement score (ie CPE). Here are the questions I have about the training:
- Based on the formula of CPE, it’s possible CPE > 1.0, especially when we use a larger constant for minimum cost. How could we treat the problem as a classification problem?
- For the labeling data for the training, are they the positive and negative examples from the history interactions?