Why do shared accounts create heterogeneous preference drift within single user profiles?
This reads the question as: when several different people use one login (a household sharing a Netflix or Amazon account), the profile fills with incompatible signals — and you want to know why that looks like 'preference drift' to a recommender rather than just noise. The corpus doesn't study shared accounts by name, but it has sharp material on the underlying problem: a single profile that contains more than one taste.
This explores why one account can hold contradictory preferences at once — the shared-household case where a recommender sees one 'user' but is really watching several. The corpus doesn't tackle shared logins head-on, but it converges on the mechanism from a more useful angle: the assumption that one profile equals one coherent taste is itself the bug. Several notes argue that even genuine individuals aren't monolithic. The AMP-CF work models each user as a *mixture* of latent personas, weighted by what's being recommended right now, rather than a single averaged taste vector Can modeling multiple user personas improve recommendation accuracy? Can attention mechanisms reveal which user taste explains each recommendation?. A shared account is just the extreme version of this: the personas belong to different people, so the 'drift' isn't one mind changing — it's the model collapsing several minds into one and watching the blend wobble.
What makes this read as *drift* rather than obvious mismatch is timing. Per-user concept-drift research shows preferences move on individual timescales for individual reasons, so population-level drift detectors miss it Why do global concept drift methods fail for recommender systems?. On a shared account, the 'individual timescale' is fake — what looks like a single person's evolving taste is actually two people interleaving sessions. A drift model dutifully tries to fit a smooth trajectory through points that were never on one curve. Related work on periodicity sharpens this: HyperBandit treats *time-of-period* as a context dimension, recovering weekly and daily cycles a change-point detector reads as instability Why do recommendation systems miss recurring user preference patterns?. A shared account often has exactly this structure — kids' content in the afternoon, adult content at night — so the 'drift' is partly a schedule, not a change of mind.
There's also a measurement layer. Ratings from the *same* person already shift by multiple stars across sessions, driven by mood, anchoring, and rating-behavior rather than stable preference Why do the same users rate items differently each time?. Pile two raters into one profile and you stack idiosyncratic noise on top of genuine taste difference — the signal stops being recoverable from the aggregate at all.
The most counterintuitive piece is what happens when the model tries to 'fix' a muddled profile by borrowing from someone similar. PRIME finds a U-shaped error curve: the *most similar* profile substitutions cause the steepest performance drops, because the system confidently applies preferences that are nearly-but-not-quite right Why do similar user profiles produce worse personalization errors?. That's the shared-account trap in miniature — a profile that's 80% one person looks coherent enough to act on confidently, and the confident wrong call hurts more than an obvious mismatch would.
The takeaway the corpus hands you, almost sideways: heterogeneous drift inside one profile isn't a data-quality accident to be cleaned away — it's evidence that 'one profile, one taste' was the wrong unit all along. The systems that handle it best (persona mixtures, per-user temporal models, time-as-context) don't denoise the profile; they stop assuming it was ever singular. If you're curious where this goes next, the same logic resurfaces in reward modeling, where personalizing too tightly to one 'user' amplifies echo chambers precisely because it removes the averaging that masked the incoherence Does personalizing reward models amplify user echo chambers?.
Sources 7 notes
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.
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.
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.
PRIME shows a U-shaped error curve where most-similar profile replacements cause steepest performance drops. The model confidently applies wrong preferences when profiles are nearly but not truly matched, an uncanny valley effect more harmful than obvious mismatch.
Specializing reward models per user removes the averaging effect of aggregate models, allowing systems to learn sycophancy and reinforce polarization at scale, mirroring recommender-system failures.