Feature engineering

How this question breaks down

1. 1Feature typesCore22 minEvery raw field reaches the model as numbers, and each kind of field (a skewed count, a category with millions of values, a paragraph of text, a timestamp) needs its own transformation to keep its signal and drop what misleads.
2. 2EmbeddingsCore19 minGive each book, author or reader a short learned vector instead of a one-hot slot among millions, so the model can share what it learns between similar IDs, and learn what that costs in memory, in cold starts and in versions that can't be mixed.
3. 3CrossesCore19 minWhen a combination of features means something neither feature does alone, a model that only adds evidence misses it. Crossing features by hand, factorising the crosses, or learning them with cross layers each let the model say "this and that".

Work through the parts in order and you’ve covered the full design. Each part is one topic.