ECHO keeps long user histories searchable for recommendation

How can a recommendation system use a person’s very long history without searching every past action for every candidate item, or losing specific actions in a compressed summary?

ECHO addresses this by retaining original records that can be searched for each candidate, while using teacher guidance to decide what to keep.

ECHO selects original user-history records with teacher guidance, then retrieves target-relevant records for ranking from that smaller set.

ECHO selects 10,000 original records from a user’s history using a teacher trained to identify target-relevant behaviors. For each candidate item, a retriever searches those records. This keeps specific past actions available while limiting the history searched during ranking.

Across eight offline tasks on two datasets, ECHO had the highest GAUC, a measure of item-ranking quality, among six comparison methods.

The result supports its use for long-history ranking in these evaluations; a separate analysis estimates lower sequence-branch computation than STCA at 100,000 events.

The evidence comes from the reported offline tasks, analytical computation estimates, and a seven-day test in one production recommendation scenario. ECHO can still discard events needed for extremely long-tail targets; its selector adapts through training and refreshes, not to each online target.