基于DeBERTaV3的Token表达与联邦学习的高校图书馆推荐系统

Federated Recommendation System for University Libraries Based on DeBERTaV3 Token Representation

摘要: 为提升高校图书馆推荐系统的个性化水平并保护用户隐私，本文提出一种融合联邦学习与Token级Transformer模型的图书推荐方法。该方法以Amazon Kindle Store数据集为基础，模拟多个高校图书馆作为联邦客户端，每个客户端本地训练模型，无需上传原始用户数据，从而有效保护读者隐私。在模型设计方面，本文采用DeBERTaV3作为Token-based Transformer骨干网络，将图书的标题与评论文本编码为多层次语义Token，实现对用户阅读行为和兴趣的深层建模。通过局部训练与中心聚合的联邦机制，构建出具备全局泛化能力的推荐模型。实验在Top-N推荐准确率、隐私保护强度与跨客户端迁移效果等维度进行了系统评估，结果表明所提出方法在保持推荐性能的同时兼顾数据安全性，适用于多校园、多用户异构环境下的个性化推荐任务。

Abstract: To enhance the personalization of university library recommendation systems while preserving user privacy, this paper proposes a book recommendation method that integrates federated learning with a Token-level Transformer model. Based on the Amazon Kindle Store dataset, the method simulates multiple university libraries as federated clients, where each client trains a local model without uploading raw user data, thus effectively safeguarding reader privacy. In terms of model architecture, DeBERTaV3 is adopted as the Token-based Transformer backbone to encode book titles and review texts into multi-level semantic tokens, enabling deep modeling of user reading behaviors and preferences. By combining local training with centralized aggregation through the federated mechanism, a globally generalizable recommendation model is constructed. Extensive experiments are conducted on Top-N recommendation accuracy, privacy preservation strength, and cross-client generalization. The results demonstrate that the proposed method achieves strong recommendation performance while maintaining data security, making it well-suited for personalized recommendation tasks in multi-campus, multi-user heterogeneous environments.