Learning trails with bounded exploration
A good feed explains why something appears and optimizes for useful action, not helpless continuation.
Current explanation
The first ranking layer should remain understandable: explicit more, less, and learned signals; topic similarity; recency; short-session fit; and a deterministic exploration term. Opening, answering, returning, and successfully explaining are better outcomes than raw dwell time. Generation stays asynchronous and editorial; ranking must not depend on a model being available.
Lesson path
- 01
Choose signals
completeRank explicit learning actions above passive context.
- 02
Add diversity
currentPrevent adjacent records from collapsing into one topic.
- 03
Observe outcomes
nextDefine what counts as useful enough to return to.
Open questions
- When is there enough outcome data to justify a contextual bandit?
Selected Q&A
Should the feed literally be infinite?
It can keep producing continuations, but it should preserve stopping points, stable URLs, reasons, and a resumable trail.
Next actions
- Compare the current deterministic score with ten hand-ranked sessions.
Sources
- TikTok recommendation overview
Public description of weighted behavior signals and diversity.
- ICAP framework
Framework distinguishing passive, active, constructive, and interactive engagement.
Revision trail
· learning:learning-trails-with-bounded-exploration@r1
Separated ranking, learning outcomes, and asynchronous generation.