> Popular songs (i.e. if almost everyone has a Beetles album in their playlist, getting "people who bought this also bought" recommendations for anything would

> Popular songs (i.e. if almost everyone has a Beetles album in their playlist, getting "people who bought this also bought" recommendations for anything would list Beetles I've been learning recommendation engines by looking at peoples' Steam games libraries. One feature of the data set is that many, many people own multiple versions of Counter-Strike as well as Team Fortress 2. So "a high number people who bought [almost any game] also bought Counter-Strike: Global Operations" is a recurring problem with a naive recommender. What I've been learning how to do is weight recommendations by how 'surprising' they are, for want of a more accurate term. If 80% of people who own Game A also own Game B, but only 5% of the total population owns Game B, then we should upweight that relationship.