Order carries meaning
Static collaborative filtering treats your history as a bag of items. Sequential recommendation instead respects the order of your actions, because what you watched last often predicts what you want next better than your all time average taste.
Modeling the sequence
- Represent each item as an embedding.
- Feed the ordered list through a sequence model such as a recurrent network or a self attention transformer.
- Read out a state that summarizes recent intent and predict the next item.
Self attention based models let any past item attend to any other, capturing both recent and long range patterns without the vanishing gradients of recurrence.
Training objective
- Slide along the sequence, predicting each next item from the prefix before it.
- This next item prediction gives many training signals per user.
Why it helps
- It captures evolving intent, like a burst of interest in cooking videos.
- It models short term context that static taste vectors smooth away.
Key idea
Sequential recommendation models the order of a user history with recurrent or self attention layers, predicting the next item to capture evolving short term intent that static taste vectors miss.