Temporal Dynamics, Recency, and Seasonality

A fashion platform retrained monthly on a one-year window. Seasonal items kept their pull long after the demand for them had ended. Then a rapid weather shift and a local festival created a pattern the model had no way to represent.

Time can affect user taste, item popularity, inventory, context, competition, and label maturity. A single timestamp feature rarely captures these mechanisms. The temporal models worth having are decay, time bins, periodic features, recency-aware neighborhoods, dynamic factors, sequence gaps, trend sources, and lifecycle states. Each has to be built point-in-time. Whoever builds one owes an answer for why that timescale and not another.

The rest of this lesson answers that question with published work rather than assertion. A drift model measured on six years of Netflix ratings. A production feature at YouTube whose omission flattens a launch curve. A public health model that learned the calendar. Four standard baselines, on four standard datasets, recommending items that did not yet exist.

Five clocks, not one. How long ago the event happened, where the calendar is, how old the item is, how far along the user is, and what the team itself changed last month.

The item-lifecycle clock is not a diagram label. It is a shipped production feature. YouTube's candidate-generation network is fed the age of every training example. The 2016 paper that describes it gives the reason: "Machine learning systems often exhibit an implicit bias towards the past because they are trained to predict future behavior from historical examples." Figure 4 shows what that bias costs on a single video. With the example-age feature the model reproduces the observed upload-time popularity spike. Without it the model "would predict approximately the average likelihood over the training window". That is a flat line drawn through a curve that is not flat. At serving time the feature is set to zero or slightly negative. That is what makes it a lifecycle clock rather than a leak.

One question underlies the row: how much should an old event still count? A hard time window answers with a cliff. Exponential or learned decay answers with a half-life somebody has to choose. A periodic or state model answers that age was the wrong variable, and that what matters is where the user sits in a cycle or a lifecycle.

The first two options are rejected in the abstract of the best-known drift model, from 2009: "Classical time-window or instance-decay approaches cannot work, as they lose too much signal when discarding data instances." What it proposes instead is the column this comparison does not have — "creating a model tracking the time changing behavior throughout the life span of the data". Reading the row as a menu of three is therefore already a choice. The fourth answer is to keep everything and let the drifting parts drift.

Future popularity leakage makes the past look easier than it was. Delayed-label bias makes the recent past look worse. Both are counted in the same evaluation, and they push the number in opposite directions.

Calendar overfit has a documented case with dates and counts attached. Google Flu Trends overshot the actual level in 2011-2012 by more than 50%. It reported overly high flu prevalence in 100 out of 108 weeks from 21 August 2011 to 1 September 2013. And it completely missed the non-seasonal 2009 influenza A-H1N1 pandemic — the episode that did not follow the calendar. Four researchers writing in Science in 2014 put the diagnosis in one line: "In short, the initial version of GFT was part flu detector, part winter detector." The errors were autocorrelated week to week, and varied in direction and magnitude with the time of year. Much of the drift they attribute to Google altering its own search algorithm underneath the model. That is the system-change clock and the periodic clock failing inside a single case.

Five clocks run underneath this lesson. The first step is only to name them: event time, availability, label maturity, lifecycle, and system change. The sequence closes by comparing predicted against observed value by age, horizon, and cohort. That comparison is what makes a temporal design reviewable rather than assumed.

Step three — compute every aggregate as known at the ranking moment — is the one most often skipped. There is a measurement of what skipping it costs. Four researchers trained BPR, NeuMF, SASRec and LightGCN on MovieLens-25M, Yelp, Amazon-music and Amazon-electronic, using the standard leave-last-out split. Their finding: "Through experiments, we show that all models indeed recommend future items that are not available at the time point of test instance, as the result of data leakage." The accuracy changes caused by including different amounts of future data reach magnitudes as high as 89.5%, with no consistent pattern. So the leak does not merely inflate a score by a known margin. It makes the ranking between models unpredictable. Point-in-time construction is not hygiene. It is the difference between a comparison and a coin toss.