Do look-alike users help more when the current session is sparse or vague?

This explores whether borrowing signal from similar users matters more when the active session is thin or ambiguous. The corpus says look-alike users are a real and useful channel — but it also warns that 'who resembles you' is a sharper knife than it looks, and pointing it at a vague session can cut the wrong way.

The clearest case for look-alikes comes from conversational recommendation, where systems that lean only on the active dialogue lose preference structure that traditional recommenders captured for years Can conversational recommenders recover lost preference signals from history?. That work argues for three channels — the current session, the user's own history, and look-alike users — but with a crucial condition: the look-alike signal must be conditioned on current intent. So look-alikes don't simply substitute for a sparse session; they fill it in *as filtered through whatever fragment of intent the session does reveal*. A vague session is exactly where this matters, but it's also where the conditioning signal is weakest — which is the tension at the heart of your question.

The danger shows up vividly in the personalization-error work: replacing a user's profile with the *most similar* available profile produces the steepest accuracy drops, a U-shaped curve where near-but-not-true matches are more harmful than obvious mismatches Why do similar user profiles produce worse personalization errors?. The model confidently applies preferences that are almost right. When a session is sparse, you have less evidence to catch that the look-alike is subtly wrong — so the uncanny-valley failure mode is precisely most likely when you'd most want to reach for a neighbor.

There's also a quieter finding that complicates the 'just retrieve similar people' instinct: across personalization tasks, recency-based recall beat similarity-based retrieval, and abstracted preference summaries beat replaying specific past interactions Does abstract preference knowledge outperform specific interaction recall?. Similarity, in other words, is not automatically the best routing signal — and a candidate-conditional view of the user, where the representation is re-weighted at prediction time against what's actually being recommended, tends to outperform a fixed similar-user lookup Can modeling multiple user personas improve recommendation accuracy?. That points the same direction as the CRS finding: look-alikes earn their keep when they're conditioned on the moment, not pasted in wholesale.

So the honest synthesis: look-alike users are most valuable when the session is sparse — that's the gap they're built to close — but sparsity is also what makes them most dangerous, because there's less in-session evidence to keep the neighbor honest. The corpus's resolution isn't 'use more similar users when you know less.' It's 'condition harder on current intent, prefer abstracted and recent signal over raw similarity, and treat the most-similar neighbor as a confident-but-risky guess.' What you didn't know you wanted to know: the closer the match, the more it can hurt.