Time-Aware Recommendation
Taste has a clock on it. What you want at 7am on a weekday is not what you want at 11pm on a Saturday, and what you wanted three years ago may embarrass you now.
Three different things time does
- Periodicity — hour of day, day of week, season. Breakfast recipes, Monday-morning playlists, December gift shopping; encoded as a context feature.
- Recency — a recent interaction says more about current intent than an old one, so interactions are weighted by an exponential decay before the model sees them.
- Drift — both users and items change; timeSVD++ is the canonical treatment, making user bias, item bias and the user factors themselves functions of time, fitted on Netflix Prize data where rating behaviour shifted measurably across years.
Why it matters
A recommender is trained on logs stretching back years and serves a request happening now. Without a time signal the model averages over all of it — offering a genre the user left behind, or treating last week's trend and a three-year-old binge as equally informative. This is not session-based recommendation: that reads the order of events inside a visit, while time-aware models read the clock value itself.
What goes wrong
Time-dependent parameters do not extrapolate — a factor fitted as a function of time has no honest value for tomorrow — and any of this evaluated on a random split will look better than it is.
Time-aware recommendation treats the timestamp as a modelling signal rather than metadata, covering three effects: periodic patterns such as hour of day or season, recency weighting that discounts older interactions, and long-term drift in user taste and item appeal, handled by time-dependent biases and factors as in timeSVD++. It keeps a model trained on years of logs from averaging across them when answering a request happening now.
A good barista knows that the person who orders a double espresso at 8am wants a chamomile tea at 9pm. Same customer, different hour, different answer.
Time-aware recommendation is software paying attention to the clock and the calendar. It notices that you watch comedies on Friday nights and documentaries on Sunday mornings, that what you clicked yesterday matters more than what you clicked in 2019, and that your tastes have genuinely moved on. Without that, a service averages your whole history into one bland picture of you — and keeps recommending the band you stopped listening to years ago.