Recommendations must account for co-viewing
Aliases: household recommender · group TV recommender
What it is
“Because you watched” on a TV home assumes the decision maker is the current profile’s one person. The sofa is often two or more people picking one title to watch together tonight. Taste overlap, the rating floor, who refuses—none of that matches a solo portrait. Recommendations must account for co-viewing means: when the situation is watching together, candidates should be ranked for the people in this room, not a single person’s Resume and private list spread as the household menu.
Why it happens
Solo recommenders optimise one person’s clicks and completions. Co-viewing utility is what several people can all accept: intersection is safer than union, and union will surface genres only the owner wants. Rating is a hard constraint—when a child is present, adult-profile matches still cannot lead the first screen. Session length differs too: watching together usually means picking one appropriately long title once, not pinning one person’s short-form or an unfinished three-season serial in first place.
The system rarely gets an explicit “this is co-viewing.” Proxy signals include a recent switch to a household profile, profiles flipping back and forth, evening hours, focus jitter from the remote changing hands. All of those are noisy. Cleaner is letting people choose “just me / household / with kids,” then freezing the profile set used for this session. When the session ends, recommendations fall back to the current profile’s solo mode, so a title someone conceded for the group is not written into their long-run portrait—otherwise a cartoon watched to keep the peace pollutes tomorrow’s alone Home.
Studying it
Co-viewing recommendation cannot be scored by solo click-through. Seat two people (or a household) at one screen to pick “what we watch tonight,” and compare a solo list, an intersection or negotiated list, and an explicit co-viewing-mode list on acceptance, vetoes, and session length.
Independent variables: how portraits merge (one person / average / strictest rating / explicit session mode), who is present, whether the list discloses “recommended for whom.” Dependent variables: share of first titles accepted, veto count, whether someone stands up to switch profile, later regret.
A “household model” trained offline on a mixed family history usually listens to whoever used the account most, which is not the people in the room tonight. A lab that instructs “you must pick one” suppresses the real-life veto and the walk-away to a phone. If “for whom” appears on screen, that itself changes disclosure and has to be logged separately from ranking quality.
Where it stops holding
Alone, headphones, a second screen in a private room: a co-viewing model is noise and should revert to solo. Live sport and news are co-presence more than co-decision; recommendation slots should yield to what is on. A one-night guest should not write clicks into the owner’s portrait, and should not be hosted with the owner’s full history—a temporary session fits that evening better than a new formal profile. Forcing the algorithm to guess who is present, a wrong guess is worse than a conservative household cover wall.
Applying it
- Offer an explicit session choice: just me, household, with kids. After that, the first screen ranks intersection under the strictest rating, and marks the list as this sitting’s watch list.
- Do not default-write playback in a co-viewing session into a member’s “because you watched.” Require confirm, or write only to a household profile.
- Once “with kids” is selected, adult genres leave the first screen; they are not merely down-ranked with half a cover still showing.
- Verify with two people of different taste, picking a title on a solo Home and in a co-viewing session. Count vetoes and requests to switch profile. If the co-viewing session still leads with one person’s Resume, the mode has not taken effect.