In lesson 5, Jeremy goes over seeing high biased people like anime watchers that rate all anime movies high.
I understood it as, high bias for a user means that the user has more error in prediction overall when predicting all types of movies since his weights were shifted towards favoring anime movies while someone of similar embedding might not so the bias soaks up error over time to counter the fact that he favors anime movies, so it offsets his predictions to make it possible for the prediction to both fit nonanime and anime movies better for that user.
Maybe I’m misunderstanding, can someone please shed light to this?
Yes, but if you can determine if the person likes anime based on their bias you can determine that they will probably rate anime higher. If you can determine that a person does not like anime you can determine that they will probably rate other types of movies higher. If I remember correctly this part of the lesson was based on the movie ratings of INDIVIDUAL users, so their individual bias is a major factor in determining how they will individually rate a movie. (colab filtering?)
The general idea is that people who like anime ONLY will have fairly similar embeddings, while people who have both will have a combination. There is a lot that can be represented in the bias though.