How do per-user concept drift and per-period periodicity combine in time-varying preferences?
This explores how two different kinds of time-based preference change—each user's tastes slowly shifting on their own schedule (drift) versus tastes that cycle on a clock (daily/weekly periodicity)—get modeled together rather than treated as one problem.
This explores how two different kinds of time-based preference change combine: per-user concept drift (each person's tastes shifting gradually, on their own timeline and for their own reasons) and per-period periodicity (tastes that recur on a clock—what you want on a weekday morning versus a Friday night). The corpus treats these as genuinely separate signals that fail when collapsed into one. Drift is about slow, irreversible change; periodicity is about predictable return. A system tuned only for drift will read every Friday as 'novel evidence' that your taste changed, when really it just came back around. The interesting move in the corpus is to stop forcing both through the same machinery. The foundational claim is that drift itself must be modeled per-user, not globally: population-level change-point detection misses the point because preferences shift on individual timescales for individual reasons, and good temporal modeling has to preserve long-term signal while discounting transient noise Why do global concept drift methods fail for recommender systems?. Periodicity then enters as a complementary mechanism rather than a competitor—HyperBandit conditions a hypernetwork on time-of-period to *generate* a user's preference parameters, so matching time periods retrieve matching preference functions instead of being treated as fresh drift to chase Why do recommendation systems miss recurring user preference patterns?. The two combine cleanly precisely because they answer different questions: drift asks 'has this person genuinely moved?', periodicity asks 'where in the cycle are we?'
What you didn't ask but the corpus quietly raises: a lot of what *looks* like drift or periodicity is actually measurement noise. The same user rates the same item differently across sessions—shifting by multiple stars—because ratings reflect rater mood, anchoring, and rating-behavior, not just stable preference Why do the same users rate items differently each time?. This is the trap underneath the whole question: before you attribute a change to drift or a cycle, you have to separate the signal from the rater's own inconsistency. The behavioral-science decomposition of annotations into genuine preferences, non-attitudes, and constructed preferences makes the same point from the alignment side—not every recorded response measures stable taste, and treating them uniformly contaminates the model Do all annotation responses measure the same underlying thing?.
There's also a representational fork worth seeing. One path says preferences are best stored as *abstract* summaries rather than replayed interactions—semantic memory beats episodic recall, and notably recency-based recall beats similarity-based retrieval, which is itself a temporal stance: what you did lately matters more than what merely resembles now Does abstract preference knowledge outperform specific interaction recall?. Another path says a user isn't one drifting vector at all but several personas, weighted dynamically by what's being recommended right now Can attention mechanisms reveal which user taste explains each recommendation? Can modeling multiple user personas improve recommendation accuracy?. That reframes both drift and periodicity as *which persona is active when*—a weekday-work persona and a weekend-explorer persona don't drift into each other, they alternate, which is periodicity by another name.
The payoff: 'time-varying preference' isn't one phenomenon. It decomposes into slow per-user drift, fast per-period cycles, noisy rating behavior, and shifting persona activation—and the systems that work are the ones that route each to its own mechanism rather than asking a single drift detector to explain all four.
Sources 7 notes
User preferences shift on individual timescales for individual reasons, making population-level drift detection ineffective. Per-user temporal modeling that preserves long-term signals while discounting transient noise is required.
HyperBandit conditions a hypernetwork on time-of-period to generate user preference parameters, capturing weekly and daily cycles that change-point detection misses. This treats time itself as a context dimension, so matching time periods retrieve matching preference functions rather than treating each period as novel evidence.
Amatriain et al. found that the same user gives substantially different ratings to the same item across sessions, shifting by multiple stars. This noise stems from temporal inconsistency, rater-specific biases, and anchoring effects—making ratings reflect both preference and rating-behavior rather than stable preference alone.
Behavioral science reveals that annotations contain genuine preferences, non-attitudes, and constructed preferences—distinguishable by consistency across measurement conditions. Treating them uniformly contaminates reward model training and downstream alignment.
PRIME framework shows semantic memory (preference summaries, parametric encodings) consistently beats episodic memory (retrieved past interactions) across models. Recency-based recall outperforms similarity-based retrieval, and task fine-tuning exceeds preference tuning methods.
AMP-CF represents each user as multiple latent personas weighted dynamically by candidate item. This makes recommendations both diverse and interpretable—each suggestion traces to the specific persona preference it satisfies—without requiring post-hoc reranking.
AMP-CF separates user representation into latent personas weighted by attention to the candidate item. This candidate-conditional approach improves accuracy by adapting the user representation at prediction time and produces inherent explanations for why items were recommended.