Summary
In this chapter, we discussed the need for different types of recommendation engines, from non-personalized ones to rating- and content-based ones, as well as hybrid models.
We learned that content-based recommendation engines use feature vectors and cosine similarity to compute similar items and similar users based on content alone. This allows us to make recommendations via k-means clustering or tree-based regression models. One important consideration is the embedding of categorical data, which, if possible, should use semantic embedding to avoid confusing similarities based on one-hot or label encodings.
Rating-based recommendations or collaborative filtering methods rely on user-item interactions, so-called ratings or feedback. While explicit feedback is the most obvious possibility for collecting user ratings through ordinal or binary scales, we need to make sure that those ratings are properly normalized.
Another possibility is to directly observe the feedback...