军事知识服务平台
一种基于子序列的隐私保护多行为序列推荐方法
A Subsequence-Based Privacy-Preserving Multi-Behavior Sequential Collaborative Recommendation Method
摘要 随着互联网的飞速发展与大数据时代的到来,推荐系统在电子商务、社交媒体和短视频平台等众多领域愈发重要,成为解决信息过载、提升用户体验的重要手段.在实际应用场景中,用户历史交互数据具有多行为性和序列性等特征,以电商领域为例,用户通常会通过浏览、收藏和购买等行为与物品产生交互.相较于只考虑单一行为类型的传统推荐方法,融合多种行为类型与序列信息的推荐方法能更全面地理解用户偏好,进而提供更贴合用户兴趣的个性化推荐.随着我国《中华人民共和国个人信息保护法》和欧盟《通用数据保护条例》等隐私法规的出台,用户隐私意识随之提升,推荐系统需在严格合规的前提下精准挖掘多行为序列中的用户偏好,而多行为序列中不同交互行为间存在复杂的时序关联,联邦多行为序列推荐(Federated Multi-Behavior Sequential Recommenda-tion,FMBSR)由此面临双重挑战:一方面,需解决多行为序列信息建模难题,以有效捕捉用户动态偏好;另一方面,FMBSR 虽可规避数据集中存储的隐私风险,但如何实现隐私保护与推荐性能的平衡,仍是其落地应用的主要障碍.为此,本文提出以可解释性强的协同过滤为核心框架的基于子序列的隐私保护协同推荐方法(Subsequence-based Privacy-preserving Collaborative Recommendation,SPCR),该方法主要包含三个模块:(1)密钥初始化:采用对称与非对称加密相结合的方式,为后续全流程隐私保护奠定基础;(2)计算相似度:设计兼具序列信息与多行为信息感知的相似度度量机制,将多行为序列信息融入用户相似度度量,提升相似度计算准确性;(3)预测评分:设计近期行为模式一致性模块,捕捉用户购买不同目标物品前的行为模式差异,动态调节相似度以适配目标物品.在JD 和 UB 两个真实数据集上的实验表明,SPCR 兼具推荐性能与隐私保护的优势.相较于无隐私保护的方法,它优于 FLAG、NextIP 等基线模型,与 GEAR 相当;相较于 FedPerGNN、LP-GCN 等隐私保护类方法,它在推荐性能上具有显著优势.在 JD 数据集上,SPCR 性能略低于 GEAR,其 Rec@10 为 0.1981、N DCG@10 为 0.1114,较次优基线 NextIP 分别提高约 10.86%和 9.42%;在 UB 数据集上,SPCR 在 Rec@10 上优于 GEAR,它的结果(Rec@10=0.0932 和N DCG@10=0.0604)较次优基线 FLAG 分别提高约 4.48%和 7.67%.同时,SPCR 与未结合隐私保护的中心化方法(SCR)在 Rec@10 和 N DCG@10 上结果完全一致,验证了隐私保护方案的无损性.SPCR 是 FMBSR 领域第一个实现多行为序列精准建模、隐私安全、性能无损与可解释性强的方法,在隐私保护、推荐性能与可解释性等方面取得了有效平衡.
Abstract:
With the development of the Internet,and the advent of the big data era,recommen-der systems have become increasingly important in numerous fields such as e-commerce,social media,and short-video platforms,and have become an important means to address information overload and enhance user experience.In practical application scenarios,users'historical interac-tion data have characteristics such as multi-behaviorality and sequentiality.Taking the e-com-merce field as an example,users usually interact with items through behaviors such as browsing,collecting,and purchasing.Compared with traditional recommendation methods that only consid-er a single behavior type,recommendation methods that integrate multi-behavior types and se-quential information can more comprehensively understand user preferences,thereby providing personalized recommendations that are more in line with users'interests.With the introduction of privacy regulations such as the"Personal Information Protection Law of the People's Republic of China"and the"European Union's General Data Protection Regulation",users'awareness of privacy has correspondingly increased.Recommender systems are required to accurately mine us-er preferences in multi-behavior sequences under the premise of strict compliance,while there are complex temporal correlations between different interactive behaviors in multi-behavior se-quences,so federated multi-behavior sequential recommendation(FMBSR)thus faces dual chal-lenges:on the one hand,it is necessary to address the difficulty of multi-behavior sequential in-formation modeling to effectively capture user preferences;on the other hand,although FMBSR can avoid the privacy risks associated with centralized data storage,how to achieve a balance be-tween privacy protection and recommendation performance remains a key obstacle to its practical application.To this end,this paper proposes a subsequence-based privacy-preserving collabora-tive recommendation method(SPCR)with a core framework of collaborative filtering featuring strong interpretability.The method consists of three main components:(1)Key Initialization:A combination of symmetric and asymmetric encryption is adopted to lay the foundation for subse-quent full-process privacy protection;(2)Similarity Calculation:A similarity measurement mechanism that is both aware of sequential information and multi-behavior information is de-signed to integrate multi-behavior sequence information into user similarity measurement and im-prove the accuracy of similarity calculation;(3)Prediction Scoring:A recent behavior pattern consistency module is designed to capture the differences in behavior patterns of users before pur-chasing different target items and dynamically adjust the similarity to adapt to the target items.Experiments on two real-world datasets,i.e.,JD and UB,demonstrate that SPCR possesses the advantage of balancing recommendation performance and privacy protection.Compared with pri-vacy-unprotected methods,it outperforms baseline models such as FLAG and NextIP and per-forms comparable with GEAR;and compared with privacy-preserving methods like FedPerGNN and LP-GCN,it achieves more superior recommendation performance.On the JD dataset,SPCR's performance is slightly lower than GEAR,with Rec@10=0.1981 and NDCG@10=0.1114,representing improvements of approximately 10.86%and 9.42%respectively compared to the second-best baseline NextIP.On the UB dataset,SPCR outperforms GEAR on Rec@10,and its results(Rec@10=0.0932 and NDCG@10=0.0604)outperform the second-best baseline FLAG by approximately 4.48%and 7.67%,respectively.Moreover,the results of our SPCR and its centralized version SCR without privacy protection are completely consistent on Rec@10 and NDCG@10,verifying the losslessness of the privacy protection scheme.Our SPCR is the first method in FMBSR with accurate multi-behavior sequential modeling,privacy security,loss-less performance,and strong interpretability,achieving an effective balance between privacy pro-tection,recommendation performance,and interpretability.
刊 名 计算机学报 ISTICEIPKU
年,卷(期): 2026, 49(9)
分类号: TP311
机标分类号: TP393.09 D923.8 G250
在线出版日期: 2026年10月8日