FM-Intent: Predicting User Session Intent with Hierarchical Multi-Task Learning

Sejoon Oh, Moumita Bhattacharya, Yesu Feng, Sudarshan Lamkhede, Ko-Jen Hsiao, Justin Basilico21 May 2025

Problem

A recommender that only predicts the next item sees what a user played, not why. The same person may want a new release in one session and the next episode of a show in another, and a next-item model has no explicit signal for that difference. Common intent models add intent heads next to the item head, but the tasks do not depend on each other. Intent is never observed directly, so it has to be built from interaction metadata.

Example from the article: At Netflix, a member who plays a follow-up episode of a show they already watch is in a "continue watching" intent, while genre preferences can shift a lot between sessions of the same member.

Solution in brief

FM-Intent runs a Transformer over the user's interaction sequence and first predicts several intent signals taken from interaction metadata. It combines them into one intent embedding with attention, then feeds that embedding into the next-item prediction. Intent is predicted before the item, so the item prediction depends on it.

Questions this case answers
  • How to infer a latent, session-level goal from implicit interaction metadata and feed it into a next-item prediction model
  • How to cluster users by current intent when no explicit intent label exists
  • How to predict the intent of a user's session and use it to improve next-item recommendations?
  • How to build user intent labels when intent is never observed directly?

Tools and methods

  • Hierarchical multi-task learning — intent heads are trained first and their output is an input to the next-item head
  • Transformer encoder — models the interaction sequence with multi-head attention
  • Attention-based intent aggregation — merges the per-signal intent predictions into one intent embedding
  • k-means++ clustering — groups users by intent embeddings
Continue reading — it's free

Google or a one-time email code. No password.