I have a classifer (one/zero labels) that was trained and hypertuned by the book. When the model was ready, I applied it to the production data: real-time and unlabeled.
After a short period (a few days); the real-time data became historic data and labeled. I checked the model real-time data (now, historic data) predictions, and analyzed them using LIME. If I see that a specific feature many times involves in FP or FN predicts and I decide to remove it.
Is it a good strategy or do I create an overfitting situation?