RecSys 2026 / Recommender Systems

Zero-Observation User Reactivation with Gap-Driven Dimensional Gating

A RecSys 2026 accepted paper defining zero-observation user reactivation and introducing DeltaGate, a frozen-backbone dimensional gating plugin for long-gap returning users.

Full publications Code

Research area

User reactivation, generative recommendation, dynamic retrieval, graph embedding, counterfactual watch-time prediction, and efficient CTR modeling form the main technical line.

Publication record

Authors
J Ding, T Liu, F Liu, H Qin, T Wu
Venue
Accepted to the 20th ACM Conference on Recommender Systems (RecSys '26)
Year
2026
Area
Recommender Systems

Problem

Returning users may have months or years without observed behavior, so a model's last user representation can become unreliable even when pre-gap history exists.

Method

DeltaGate keeps the sequential recommendation backbone frozen and learns a dimension-wise gate that routes each representation dimension between personalized history and a learned global prior.

Findings

Under the Gap-Synthesize Protocol on Amazon datasets, standard sequential recommenders decline as inactivity gaps widen; in the >365d Video Games bucket, DG-BERT4Rec reaches 0.046 Hit@10 versus 0.025 for BERT4Rec.

Deployment boundary

End-to-end retraining can reach higher absolute accuracy, while the frozen DeltaGate plugin preserves zero backbone drift, uses 66K trainable parameters, and keeps the adaptation behavior inspectable.