Loading
How a system learns what each user will like from what users did before. The course covers neighbourhood methods, matrix factorisation fitted with SGD and ALS, implicit feedback and BPR, two-tower retrieval, and the ranking metrics used to judge a list of recommendations.
01
Collaborative filtering predicts a user's taste from the behaviour of similar users or similar items. Item-based methods compare the columns of the user and item matrix.
Practise it
02
Each user and item gets a short vector, and their dot product predicts the rating. Training uses only the observed ratings, with a penalty on the size of the vectors.
03
Fix the item vectors and each user vector has a closed-form ridge regression solution, then swap. Each half step is a set of small linear solves.
Practise it
04
Clicks and views say what a user did, not how much they liked it. Bayesian personalised ranking trains the model to score a seen item above an unseen one.
Practise it
05
A user tower and an item tower map features to vectors, and a dot product scores the pair. In-batch negatives make training cheap, and a nearest-neighbour index makes serving fast.
06
A recommender returns a list, so it is scored on the list. Precision and recall at k count hits in the top k, and NDCG also rewards hits near the top.