Predicting preferences.
WHAT THE APPROACHES ARE
Content-based: similarity between items
Collaborative: patterns across users
Hybrid
WHAT MATRIX FACTORISATION DOES
Represents users and items as vectors whose product predicts preference.
WHAT THAT PROVIDES
Learning from interaction patterns without describing items.
WHAT IMPLICIT FEEDBACK IS
Behaviour rather than ratings: views, purchases, time spent.
WHY IT DOMINATES PRACTICE
Explicit ratings are rare, and behaviour is abundant.
WHAT MAKES IT DIFFICULT
Absence of interaction does not mean dislike.
WHAT THAT REQUIRES
Treating unobserved items as unknown rather than negative.
WHAT THE COLD START PROBLEM IS
No history for a new user or item.
WHAT ADDRESSES IT
Content features, popularity, or asking directly.
WHAT TO EVALUATE WITH
Ranking metrics, not error metrics.
WHY
What matters is what appears at the top, not predicted values.
WHAT TO MEASURE BEYOND ACCURACY
Coverage, diversity, and novelty.
WHY
A system recommending only popular items scores well and serves nobody.
WHAT TO TEST ONLINE
Everything, since offline evaluation cannot see responses to unseen recommendations.
