Exclude casting roles from crew matching
A casting office works across a studio’s whole slate, so two shows sharing a casting director says nothing about whether they are alike. Those roles are removed before crew is compared, rather than down-weighted.
Context
The Boys and Grey’s Anatomy came back as similar shows.
They share four crew members. All four are casting directors, all four from the same office: a superhero satire tied to a medical drama by nothing more than the people who hired their actors.
That is what sent me into the data model, and the pattern was general. A casting office works across many of a studio’s shows, so a shared casting director says nothing about whether two shows resemble each other. Casting turned out to be the second most common source of these false matches, behind only executive producers.
Decision
Remove casting roles from both shows before comparing their crew, so a shared casting director never counts toward two shows being similar.
The ingest keeps an explicit list of casting job titles to exclude rather than matching on the word “casting,” so the list is easy to check by hand. TMDb occasionally adds a new title that slips the list – a later review caught eight, including “Extras Casting” and “Location Casting,” which the list now carries.
I considered lowering casting’s weight instead of removing it. It does not work. A casting director credited on 60 episodes still says nothing about the show, so the episode weighting in ADR-04 cannot fix this one. It is a separate problem and it needs a separate answer.
After Action Review
A show’s similar list is now built from the people who actually made both shows, not from the back-office staff a studio reuses across unrelated titles. The obviously wrong matches disappeared and genuine ones took their place – shows that really do share a creator, a director, and a composer.
The cost is that a couple of shows whose only link was a shared casting office now show no match at all. That is the honest answer, and a blank is better than a wrong one.
The same exclusion carries into the full weighted recommender, ADR-04.