Problem

Dynamic recommendation retrieval must keep up with evolving user interactions, and a key signal is the short-term co-occurrence pattern that reflects current preference. Existing ways of capturing this signal face scalability and cost constraints.

Approach

RPE4Rec is a lightweight plug-in that applies relative position encoding to dynamic node retrieval. It uses a hashtable-based neighborhood memory to capture short-term co-occurrence signals with low space and time complexity, and attaches to mainstream recommendation models.

Public evidence

Experiments across multiple datasets report consistent retrieval improvements in mainstream models with minimal additional cost. The abstract does not give numeric metrics.

Takeaway

Efficient position-aware retrieval is one mechanism for keeping retrieval aligned with changing user and item behavior, the recurring condition in the adaptive-recommendation line.

Bibliographic details

Authors

Ke Cheng, Heng Chang, Pengyang Wang, Liang Gu, Jiandong Ding, Yi Cao, Junchen Ye, Bowen Du

Venue

WSDM 2026

Status

Published