Quantity isn't always quality
Yesterday I found out that TMDb splits credits into two endpoints. Today I ingested aggregate_credits, which rolls up every episode credit and returns an episode count per person.
The catalog went from 728 cast rows to 128,541 and from 1,665 crew rows to 24,482. Directors were invisible before because TV shows credit directors per episode. I had 13 director rows across 100 shows. I now have 3,270, and 480 directors work on two or more shows. Félix Alcalá directed both Breaking Bad and CSI. John Dahl directed Breaking Bad, Dexter, and Outlander.
Observation
Today the recommender got worse. 84% of my cast rows are people who appeared in exactly one episode. When I count every shared person equally, Grey’s Anatomy now returns NCIS as its top match because they share 715 guest actors. That connection means nothing. Two long-running network shows just return the same pool of working actors. More data is not better data until I weight it. I expected this because I chose a simple recommender on Day 1.
Next Up
Tomorrow I weight each shared person by how much of the show they were part of. Walter White counts as 62 of 62 episodes. A one-episode guest counts as 1 of 62. I tested the math and CSI drops from second place to eighth while Better Call Saul stays first.
See you tomorrow for Day 4!
For a more in-depth explanation, read this ADR.