Quick note: Collaborative Filtering vs. Swing

You probably have heard of Collaborative Filtering, which is to find similarity of two items by counting  the number of shared interested users. They are matched via a cosine similarity (or dot product in another word).

There comes a problem: two users are within the same interest group, the shared items may not be that similar.

Swing takes one more step: assign a weight for each pair of shared users - if the two users have shared a lot of items, let the weight of the pair of the user be smaller:  1 / (a + overlap(u1, u2) . (a is a positive term to avoid dividing by zero). 

Collaborative Filtering may also happen in similarity of users. Recommendation is made from similar users' list of items. To avoid popular item from appearing in every recommendation, weight of item is reduced by the popularity with 1/ log(1+ num_users_like_the_item)

Reference:

https://www.youtube.com/watch?v=DUUMNTDuJ3Q

https://www.youtube.com/watch?v=7O9zFMNdrZ8