When there is no profile
Many visitors are anonymous or new, so there is no long history to draw on. Session based recommendation works from only the clicks in the current session, predicting the next item from this short, fresh sequence.
Why it differs from sequential
- There is no stable user identity, just an in session click stream.
- Sessions are short, so the model must act fast on a few signals.
- Intent can shift within minutes, so recency dominates.
Common approaches
- Recurrent networks over the session clicks, the classic GRU based recommender.
- Graph models that turn the session into a small graph of item transitions.
- Nearest neighbor baselines that match the current session to similar past sessions, which are surprisingly strong.
Practical concerns
- Handle the first click of a session, where context is almost empty.
- Weight recent clicks more, since the latest action reveals current intent.
Key idea
Session based recommendation predicts the next item from only the current anonymous click stream, leaning on recency and session level patterns rather than any long term user profile.