Is it so hard to believe that they are manually tagged?
They only have several thousand videos. Each one has some sort of process that it goes through to be included in their service. Someone decides which screenshots will be shown, edits the intro video, writes the synopsis, etc.
Don't know if it's what Netflix actually uses, but you can use perceptual hashing[1] to identify spans that are common to multiple episodes of the same series.
Is it naive to think that just as the shiws provide previews, box art etc they would provide the length of their intro? Then skip intro just needs to increment the video time by that amount
In addition to the other suggestion, I'd expect it to be something like this [0] or [1], a mix of data on what people skip in the first 10 minutes and the superstar of modern machine learning: human annotations. I've seen several shows where all episodes where consistently off by a few seconds, that might be either algorithmical bias or due to some annotators personal preference.
In the case of Netflix created content. It can be as simple as a number of seconds in metadata. Things get easy if you control both the production and consumption ends.
They only have several thousand videos. Each one has some sort of process that it goes through to be included in their service. Someone decides which screenshots will be shown, edits the intro video, writes the synopsis, etc.