If you already know exactly the 'age-out' behaviour that you want to capture, then maybe the best thing to do is bolt that on after doing your statistical inference.
This keeps the 'complicated' bit of your system that uses the statistical models simpler.
Also, you'll know exactly what the heuristic does, and be able to predictably hack changes into it in a single place, without affecting your inference code.
If you instead put the aging-out into your objective function, you'll have a more elegant implementation, but the 'age-out' behaviour might surprise you.
Well-specified models are great, but heuristic hacks that force the system to behave how you want also have a place. In general, I would start out by only using the statistical modelling in places where you can't describe exactly what you want with a heuristic.
What the best approach actually is depends, and I'd need to be implementing the system to say more.
If you already know exactly the 'age-out' behaviour that you want to capture, then maybe the best thing to do is bolt that on after doing your statistical inference.
This keeps the 'complicated' bit of your system that uses the statistical models simpler.
Also, you'll know exactly what the heuristic does, and be able to predictably hack changes into it in a single place, without affecting your inference code.
If you instead put the aging-out into your objective function, you'll have a more elegant implementation, but the 'age-out' behaviour might surprise you.
Well-specified models are great, but heuristic hacks that force the system to behave how you want also have a place. In general, I would start out by only using the statistical modelling in places where you can't describe exactly what you want with a heuristic.
What the best approach actually is depends, and I'd need to be implementing the system to say more.