Time Based Content Based Filtering

Let’s say we want to incorporate the date the rating was made in order to reflect user preferences changing over time. What would be a good way to incorporate that input? In my mind, it neither fits nicely into the user or movie inputs. I’d be tempted to add a third network for the date, taking the dot product of all three before offering a prediction, but that also doesn’t sit quite right in my mind.

I do not think a 3rd network is a good idea.

Time is a difficult concept as a feature, because there are lots of time parameters. Elapsed time, time of day, day of week, month, season, year, etc.

Yeah, I agree, that’s what I’m finding so tough thinking about it. I was thinking of a feature along the lines of seconds since the rating had been made, with 0 as the input when making a prediction. Maybe that makes sense as part of the user network? What are your thoughts?

Using 0 for the time when making a prediction would not be wise, because if time is a linear feature, you’d be asking for a prediction that is earlier than any of the training examples.