Collaborative Filtering
Whenever we discuss about collaborative filtering, the following image is implicitly made as a anchor point to talk about how collaborative filtering discovers ‘similar users’ and use that to recommend unseen items.
$$ \begin{array}{c|cccccc|} & i_1 & i_2 & & \cdots & & i_n \\ \hline u_1 & & \checkmark & & & & \checkmark \\ %\hline u_2 & \checkmark & \checkmark & & & & \\ %\hline & \checkmark & & & \checkmark & \checkmark & \checkmark \\ %\hline \vdots& & & \checkmark & & & \\ %\hline & \checkmark & \checkmark & & \checkmark & \checkmark & \checkmark \\ %\hline u_m & & \checkmark & & \checkmark & & \\ \hline \end{array} $$
It’s a helpful mental model and historically accurate for early recommender systems.
However, that similar users picture is only a small part of how modern recommendation models actually work. Real-world systems rarely operate directly on this matrix, and the notion of similarity today is far more nuanced than simple row-to-row comparison.
This 5 part series is a walkthrough of how collaborative filtering is practiced today. We will talk about how the row-to-row comparison idea evolves into one of the cornerstones of modern recommender systems.