Can in-session recommendation and long-horizon per-user drift be modeled in the same framework?
This explores whether the fast loop of adapting to what a user is doing right now (within a single session) and the slow loop of tracking how each person's tastes shift over weeks and months can live inside one recommendation model — or whether they're fundamentally different problems that resist a shared framework.
This question is really asking whether two different clocks can share one mechanism: the seconds-to-minutes clock of in-session signals, and the weeks-to-months clock of per-user preference drift. The corpus doesn't offer a single paper that unifies both, but read laterally it sketches why a shared framework is hard — and what the seam between the two looks like.
Start with the in-session side. Netflix's work shows that real-time, mid-session adaptation genuinely improves ranking, but at an irreducible cost: signals that arrive during a session can't be precomputed, forcing runtime recomputation, more calls, timeout risk, and bugs that are hard to reproduce How can real-time recommendations stay responsive and reproducible?. So the in-session loop is defined by latency and freshness pressure. The long-horizon side has the opposite character: it argues that drift must be modeled *per user*, because preferences shift on individual timescales for individual reasons, and population-level drift detection misses this entirely Why do global concept drift methods fail for recommender systems?. One loop is about reacting fast; the other is about remembering slowly and selectively.
The most interesting hint that these can converge comes from reframing *time itself as a context dimension*. HyperBandit conditions a hypernetwork on time-of-period to regenerate a user's preference parameters, so matching time periods retrieve matching preference functions rather than treating each moment as novel evidence Why do recommendation systems miss recurring user preference patterns?. That's a quiet unification: if 'now in the session' and 'this point in the user's weekly cycle' are both just coordinates the model conditions on, then short-horizon and long-horizon adaptation become the same operation at different scales. The cost is that you've turned a memory problem into a conditioning problem.
Another route through the corpus is structural separation rather than fusion. DEGC handles streaming recommendation by isolating parameters per task — preserving older patterns exactly while letting new parameters absorb emerging preferences, giving explicit control over the stability-vs-plasticity trade-off that replay and distillation can't match Can model isolation solve streaming recommendation better than replay?. That's effectively two timescales in one model, but kept in separate compartments rather than blended. The persona work points the same direction from the representation side: representing a user as multiple attention-weighted personas, dynamically weighted by the candidate item, lets the model surface different facets of a person at prediction time Can attention mechanisms reveal which user taste explains each recommendation? Can modeling multiple user personas improve recommendation accuracy? — which is one way a slowly-learned long-term profile and a fast in-session signal could coexist (stable personas, fast re-weighting).
So the honest answer the corpus supports: yes, but the unification is architectural, not free. The shared framework either (a) treats time as just another context coordinate so both clocks run through the same conditioning machinery, or (b) keeps stable long-term structure and fast in-session adaptation in separate compartments that interact at prediction time. What no note here resolves is the hard middle — reconciling the latency budget of in-session recomputation How can real-time recommendations stay responsive and reproducible? with the per-user, individually-paced memory that long-horizon drift demands Why do global concept drift methods fail for recommender systems?. That tension, not the modeling abstraction, is where a unified framework actually gets paid for. A nearby idea worth pulling on: framing the whole thing as one policy that decides what to do and when, rather than stitched-together components Can unified policy learning improve conversational recommender systems?.
Sources 7 notes
Netflix's in-session adaptation improves ranking by 6% relative, but precomputing is impossible when signals arrive mid-session. This forces runtime recomputation, increasing call volume, timeout risk, and making bugs harder to reproduce.
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.
DEGC uses per-task parameter isolation to handle streaming recommendation, providing explicit stability-plasticity trade-offs that experience replay and knowledge distillation methods cannot match. This approach preserves older patterns exactly while allowing new parameters to capture emerging preferences.
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.
Research shows that formulating attribute-asking, item-recommending, and timing decisions as a single graph-based RL policy achieves better joint optimization than isolated components. Separation prevents gradient signals from informing one another and fails to optimize conversation trajectory holistically.