Suggesting what someone might want.
WHAT APPROACHES EXIST
Content-based: recommending items similar to those a user liked
Collaborative: recommending what similar users liked
Hybrid, combining them
WHAT COLLABORATIVE FILTERING EXPLOITS
Patterns across users, without needing to understand the items.
WHAT THE COLD START PROBLEM IS
Having no history for a new user or a new item.
WHAT ADDRESSES IT
Content-based methods, popularity, or asking the user directly.
WHAT POPULARITY BIAS IS
Popular items being recommended more, becoming more popular still.
WHAT THAT CAUSES
Narrowing of what anyone sees, and new items never surfacing.
WHAT ADDRESSES IT
Deliberate exploration, showing some items outside the prediction.
WHAT FILTER BUBBLES ARE
Users shown progressively narrower content matching inferred preferences.
WHY THAT IS A PROBLEM
It reduces discovery and can entrench harmful patterns.
WHAT TO MEASURE
Not only engagement, but diversity and long-term satisfaction.
WHY
Optimising engagement alone produces systems that people use compulsively and resent.
WHAT TO PROVIDE USERS
Visibility of why something was recommended A way to say it was unwanted A way to see content outside the recommendation
