Dynamic User Modeling and Lightweight Knowledge Signals for Candidate Reranking: A Reproducible Two-Stage Study

Abstract

This study investigates whether introducing dynamic user information and simple knowledge signals can improve the re-ranking effect of the fixed candidate list in the recommendation system when computing resources are limited. To this end, we have constructed a reproducible two-stage recommendation framework: the first stage uses an ID-based collaborative retrieval model to generate the candidate item list, and the second stage utilizes a lightweight model to re-sort these candidates, combined with dynamic user modeling as well as semantic and knowledge-related features. We conducted experiments on the MovieLens-1M dataset and adopted a data setting based on user time sequence, evaluating the sorting ability of the second-stage model on the fixed candidate set. The experimental results show that, compared with the lightweight static ID re-ranking baseline, the dynamic re-ranking model has improved in terms of Recall@10, NDCG@10, and MRR@10 indicators. The ablation results indicate that adding more model components does not necessarily lead to better generalization ability. Some simplified configurations perform better on the test set, and the contribution of dynamic user modeling is positive but relatively limited. Further analysis also reveals that this method still faces significant difficulties in recommending rare items, but the efficiency evaluation shows that the entire process is feasible under moderate hardware conditions. Overall, this paper proposes and validates a two-stage candidate re-ranking framework, and through experimental analysis, demonstrates the role of different components in the fixed candidate recommendation scenario.