Abstract: The long tail phenomenon has long been present in sequential recommender systems, including both long-tail users and long-tail items. Although there have been many studies aiming at alleviating the long tail issue in sequential recom-mender systems, most of them only focus on either long-tail users or long-tail items unilaterally. However, the issues of both long-tail users and long-tail items frequently coexist. Neglecting one aspect will lead to poor recommendation performance in the other. Additionally, the problem of information scarcity for long-tail users and long-tail items has not been adequately addressed. This paper proposes a long-tail recommendation model utilizing gate recurrent unit dual-branch information collaboration enhancement to jointly alleviate the long-tail issues from the perspectives of both users and items, while enriching long-tail information via information collaboration enhancement. The model consists of two branches, the one for long-tail users and the other for long-tail items. Each branch is responsible for processing its respective information and is trained together to enrich the information of the other branch. Furthermore, the model introduces a preference mechanism that dynamically adjusts user preference and item popularity by calculating the influence factors for users and items, which further alleviate the problem of insufficient information in long-tail recommendation. Experiments are conducted on 6 real datasets from the Amazon series and the proposed model (LT-GRU) is compared with 6 classical models. Compared with the best results achieved by existing long-tail recommendation models, LT-GRU achieves an average improvement of 2.49% and 3.80% in terms of HR and NDCG metrics, respectively. This indicates that this work effectively mitigates the issues of the long-tail user and long-tail item without sacrificing the recommendation performance for mainstream users and popular items.

Key words: recommender system, long-tail recommendation, information collaboration enhancement, gated recurrent unit (GRU)

摘要： 长尾现象在序列推荐系统中长期存在，包括长尾用户和长尾项目两个方面。虽然现有许多研究缓解了序列推荐系统中的长尾问题，但大部分只是单方面地关注长尾用户或长尾项目。然而，长尾用户和长尾项目问题常常同时存在，只考虑其中一方会导致另一方性能不佳，且未关注到长尾用户、长尾项目各自的信息匮乏问题。提出一种利用 GRU双分支信息协同增强的长尾推荐模型（long-tail recommendation model utilizing gated recurrent unit dual-branch information collaboration enhancement, LT-GRU），从用户与项目两个方面共同缓解长尾问题，并通过协同增强的方式丰富长尾信息。该模型由长尾用户和长尾项目双分支组成，每个分支分别负责各自的信息处理，并相互训练以充实另一方的信息。同时，引入一种偏好机制，通过演算用户与项目的影响因子，以动态调整用户偏好与项目热度，进一步缓解长尾推荐中信息不足问题。在Amazon系列的6个真实数据集上与6种经典模型进行实验对比，相较于长尾推荐模型中最优的结果，所提模型LT-GRU在HR与NDCG两个指标上分别平均提高2.49%、3.80%。这表明，在不牺牲头部用户和热门项目推荐性能的情况下，有效地缓解了长尾用户和长尾项目问题。

关键词: 推荐系统, 长尾推荐, 信息协同增强, 门控循环单元（GRU）