We are having a month on month data received for training, end of every month snapshot is used. XGboost model (binary classification) is used to perform well with one month in train test and not in live, we added additional month data but the recall dropped. Is there a way to influence the model to give more importance/weight to recent data while making using previous month data to support the learning and volume.
I tried training with different month data on its own. As the data is highly imbalanced tried SMOTE ENN, tomek, but using SMOTE reduced the recall a lot.