Background: I have an app data (impressions, user activities) that I can use as features for a multiclass classifier (5 classes).
I just want to discuss about some things that our team is having a hard time finalizing:
Regarding the use of a decay factor for features—specifically, whether it makes sense to assign more weight to recent activities such as likes, comments, and shares. How would incorporating a decay factor impact the model's performance, and is it a strategy we should consider?
Our target variable is determined by a date referred to as the "target_date," which was when users were assigned a score of 1 to 5. I'm seeking input on whether it's reasonable to restrict the training data for the model to one year's worth of historical data per user or if we should utilize all available historical data, acknowledging that older users might have substantially more data than recent joiners.
Really confused especially with the first point because I cannot come to terms with the concept of adding a decay in the feature set to create the model.
Thanks in advance.