基于条件变分自编码器的联邦推荐系统冷启动问题研究

doi:10.3969/j.issn.1000-5641.2025.06.007

摘要/Abstract

摘要：

推荐系统的冷启动问题影响网站的推荐质量和用户体验, 同时还存在隐私安全问题, 这已成为该领域最具挑战性的研究热点之一. 针对该问题, 提出了一种集成联邦学习框架的条件变分自编码器模型(FedCVAE), 通过在客户端训练各自的条件自编码器模型, 学习用户、商品及用户交互序列的嵌入表示, 作为骨干推荐模型的输入, FedCVAE在服务器端完成全局模型的聚合更新并下发至客户端, 以指导本地条件自编码器模型更新超参数, 在提高模型处理稀疏数据的推荐准确度的同时, 增强模型的隐私保护能力, 从而有效缓解了推荐系统冷启动问题. 实验结果表明, 在3类典型冷启动场景中, 相较于主流推荐算法, 所提出模型的平均绝对误差指标分别降低了0.8% ~ 5.5%, Hit@5指标分别提升了1.2% ~ 5.7%, 均表现出优越的性能, 实现了更高质量的、兼具个性化体验和隐私保护增强的推荐服务.

关键词: 推荐系统, 冷启动, 联邦学习, 元学习, 变分自编码器

Abstract:

The cold-start problem of recommendation systems which affects recommendation quality, service experience, privacy, and security, has become one of the most challenging research hotspots in the field. Thus, our study proposed an integrated federated learning framework for conditional variational autoencoders (CVAE), termed as FedCVAE. Individual CVAE models were trained on each client’s local data to generate embeddings of users, items, and user interaction sequences. These embeddings were used as inputs for the essential recommendation model. The model’s global parameters were aggregated and updated at the server-side, which was subsequently disseminated back to the client-side to support local CVAE models in updating hyperparameters. While the model improves its accuracy in handling sparse data, it also enhances its ability to preserve privacy, thus effectively mitigating the cold-start problem. The experimental results indicate that in three typical cold-start scenarios, the model presented in this paper outperformed mainstream recommendation algorithms. The mean absolute error metric reduces by approximately 0.8% ~ 5.5%, and the Hit@5 metric improves by approximately 1.2% ~ 5.7%. The model demonstrated superior performance, delivering high-quality recommendation services that balance personalized experiences and enhanced privacy protection.

Key words: recommendation systems, cold-start, federated learning, meta-learning, variational autoencoder

中图分类号:

TP391

吕欣樾, 黄新力. 基于条件变分自编码器的联邦推荐系统冷启动问题研究[J]. 华东师范大学学报（自然科学版）, 2025, 2025(6): 53-62.

Xinyue LYU, Xinli HUANG. Cold-start problem of federated recommendation systems based on conditional variational autoencoder[J]. J* E* C* N* U* N* S*, 2025, 2025(6): 53-62.

图/表 6

图1

图2

表1

表2

表3

表4

参考文献 22

1	BASU C, HIRSH H, COHEN W, et al. Recommendation as classification [C]// Proceedings of the Fifteenth National/Tenth Conference on Artificial Intelligence/Innovative Applications of Artificial Intelligence. ACM, 1998: 714-720.
2	WU C H, WU F Z, LYU L J, et al.. A federated graph neural network framework for privacy-preserving personalization. Nature Communications, 2022, 13 (1): 3091.
3	HU Y F, KOREN Y, VOLINSKY C. Collaborative filtering for implicit feedback datasets [C]// 2008 8th IEEE International Conference on Data Mining. IEEE, 2008: 263–272.
4	VOLKOVS M, YU G W, POUTANEN T, et al. DropoutNet: Addressing cold start in recommender systems [C]// Proceedings of the 31st International Conference on Neural Information Processing Systems. ACM, 2017: 4964-4973.
5	KANG W C, MCAULEY J. Self-attentive sequential recommendation [C]// 2018 IEEE International Conference on Data Mining (ICDM). IEEE, 2018: 197-206.
6	QIAN T Y, LIANG Y L, LI Q, et al.. Attribute graph neural networks for strict cold start recommendation. IEEE Transactions on Knowledge and Data Engineering, 2022, 34 (8): 3597- 3610.
7	ZHAO X, REN Y, DU Y, et al. Improving item cold-start recommendation via model-agnostic conditional variational autoencoder [C]// Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, 2022: 2595-2600.
8	LEE H, IM J, JANG S, et al. MeLU: Meta-learned user preference estimator for cold-start recommendation [C]// Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. ACM, 2019: 1073-1082.
9	WANG J L, DING K Z, CAVERLEE J, et al. Sequential recommendation for cold-start users with meta transitional learning [C]// Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, 2021: 1783-1787.
10	WEI T X, WU Z W, LI R R, et al. Fast adaptation for cold-start collaborative filtering with meta-learning [C]// 2020 IEEE International Conference on Data Mining (ICDM). IEEE, 2020: 661–670.
11	WU Z C, ZHOU X. M2EU: Meta learning for cold-start recommendation via enhancing user preference estimation [C]// Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, 2023: 1158-1167.
12	MCMAHAN H B, MOORE E, RAMAGE D, et al. Communication-efficient learning of deep networks from decentralized data [EB/OL]. (2016-02-17)[2024-02-01]. https://arxiv.org/abs/1602.05629.
13	AMMAD-UD-DIN M, IVANNIKOVA E, KHAN S A, et al. Federated collaborative filtering for privacy-preserving personalized recommendation system [EB/OL]. (2019-01-29)[2024-02-01]. https://arxiv.org/abs/1901.09888.
14	LIN G Y, LIANG F, PAN W K, et al.. FedRec: Federated recommendation with explicit feedback. IEEE Intelligent Systems, 2021, 36 (5): 21- 30.
15	CHAI D, WANG L Y, CHEN K, et al.. Secure federated matrix factorization. IEEE Intelligent Systems, 2021, 36 (5): 11- 20.
16	LIN Y J, REN P J, CHEN Z M, et al. Meta matrix factorization for federated rating predictions [C]// Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, 2020: 981-990.
17	SOHN K, YAN X C, LEE H, et al. Learning structured output representation using deep conditional generative models [C]// Proceedings of the 29th International Conference on Neural Information Processing Systems. ACM, 2015: 3483-3491.
18	TIAN R Z, MAO Y Y, ZHANG R C. Learning VAE-LDA models with rounded reparameterization trick [C]// Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2020: 1315–1325.
19	SUN F, LIU J, WU J, et al. BERT4Rec: Sequential recommendation with bidirectional encoder representations from transformer [C]// Proceedings of the 28th ACM International Conference on Information and Knowledge Management. ACM, 2019: 1441-1450.
20	XIE X, SUN F, LIU Z Y, et al. Contrastive learning for sequential recommendation [C]// 2022 IEEE 38th International Conference on Data Engineering (ICDE). IEEE, 2022: 1259–1273.
21	CHONG L, LIU X Y, ZHENG R Q, et al. CT4Rec: Simple yet effective consistency training for sequential recommendation [C]// Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. ACM, 2023: 3901-3913.
22	WU C H, WU F Z, QI T, et al. FedCL: Federated contrastive learning for privacy-preserving recommendation [EB/OL]. (2022-04-21)[2024-02-01]. https://arxiv.org/abs/2204.09850.

模型	MAE				Hit@5
模型	warm	user cold	item cold	user item	warm	user cold	item cold	user item
CL4SRec	0.811	0.837	0.879	0.866	0.801	0.783	0.772	0.769
CT4Rec	0.776	0.789	0.836	0.823	0.864	0.856	0.827	0.816
FedCL	0.764	0.787	0.854	0.815	0.857	0.869	0.839	0.832
MetaTL	0.782	0.792	0.793	0.799	0.876	0.867	0.845	0.829
MetaCF	0.755	0.761	0.801	0.798	0.851	0.862	0.855	0.839
M2EU	0.745	0.759	0.798	0.793	0.849	0.863	0.852	0.843
FC-B	0.749	0.752	0.752	0.756	0.889	0.879	0.892	0.859
FC-S	0.735	0.747	0.749	0.759	0.875	0.868	0.859	0.855

模型	MAE				Hit@5
模型	warm	user cold	item cold	user item	warm	user cold	item cold	user item
CL4SRec	0.954	1.095	1.045	1.012	0.635	0.631	0.621	0.617
CT4Rec	0.932	1.054	1.023	0.967	0.705	0.669	0.643	0.623
FedCL	0.902	1.065	1.027	0.997	0.698	0.678	0.654	0.625
MetaTL	0.899	1.094	1.098	1.073	0.624	0.659	0.629	0.597
MetaCF	0.928	0.987	0.997	0.989	0.698	0.693	0.653	0.624
M2EU	0.913	1.104	0.985	0.968	0.704	0.685	0.697	0.647
FC-B	0.878	0.956	0.943	0.948	0.725	0.701	0.731	0.679
FC-S	0.866	0.962	0.934	0.957	0.732	0.712	0.729	0.654

模型		MAE				Hit@5
模型		warm	user cold	item cold	user item	warm	user cold	item cold	user item
SASRec	FC	0.735	0.747	0.749	0.759	0.875	0.868	0.859	0.855
	FC-M	0.749	0.753	0.758	0.762	0.864	0.859	0.845	0.847
	FC-C	0.764	0.767	0.771	0.783	0.849	0.845	0.832	0.839
	FC-CM	0.776	0.789	0.809	0.794	0.831	0.823	0.798	0.825
BERT4Rec	FC	0.749	0.753	0.752	0.759	0.889	0.879	0.892	0.855
	FC-M	0.756	0.769	0.763	0.774	0.878	0.867	0.876	0.843
	FC-C	0.771	0.786	0.792	0.799	0.853	0.834	0.853	0.817
	FC-CM	0.793	0.801	0.823	0.815	0.827	0.819	0.813	0.786

模型		MAE				Hit@5
模型		warm	user cold	item cold	user item	warm	user cold	item cold	user item
SASRec	FC	0.866	0.962	0.934	0.957	0.732	0.712	0.729	0.654
	FC-M	0.891	1.012	1.014	1.001	0.702	0.689	0.679	0.623
	FC-C	0.956	1.098	1.103	1.089	0.656	0.644	0.624	0.587
	FC-CM	0.998	1.112	1.148	1.119	0.602	0.613	0.603	0.567
BERT4Rec	FC	0.878	0.956	0.943	0.948	0.725	0.701	0.731	0.679
	FC-M	0.897	1.004	1.001	0.997	0.684	0.678	0.681	0.634
	FC-C	0.957	1.107	1.102	1.083	0.613	0.649	0.611	0.602
	FC-CM	1.017	1.127	1.129	1.103	0.599	0.624	0.586	0.578

[1]	冯路平, 施力业, 吴雯, 郑骏, 胡文心, 郑巍. 基于偏好级别陌生人信息辅助的推荐系统[J]. 华东师范大学学报（自然科学版）, 2024, 2024(2): 53-64.
[2]	陈道佳, 陈志云. 分层级描述感知的个性化推荐系统[J]. 华东师范大学学报（自然科学版）, 2023, 2023(6): 73-84.
[3]	高祥云, 孟丹, 罗明凯, 王俊, 张丽平, 孔超. 支持隐私保护的端云协同训练[J]. 华东师范大学学报（自然科学版）, 2023, 2023(5): 77-89.
[4]	史洪玮, 洪道诚, 施连敏, 杨迎尧. 异构编码联邦学习[J]. 华东师范大学学报（自然科学版）, 2023, 2023(5): 110-121.
[5]	朱思涵, 浦剑. 基于序列特征的点击率预测模型[J]. 华东师范大学学报（自然科学版）, 2020, 2020(4): 134-146.
[6]	王凯, 李博涵, 万朔, 张安曼, 关东海. 基于反向最远邻的商品推荐算法研究[J]. 华东师范大学学报(自然科学版), 2019, 2019(3): 63-77.
[7]	张巍, 郑骏, 逄娇娜, 白玥. 基于自相似矩阵的协同过滤推荐算法[J]. 华东师范大学学报(自然科学版), 2018, 2018(4): 120-128,146.
[8]	宋春芝, 董晓蕾, 曹珍富. 高效可验证的隐私保护推荐系统[J]. 华东师范大学学报(自然科学版), 2018, 2018(2): 41-51,62.
[9]	余文喆, 张蓉, 王立. 电子商务中的商品推荐系统[J]. 华东师范大学学报(自然科学版), 2013, 2013(3): 46-53.
[10]	宋乐怡, 熊辉, 张蓉. 下一代移动推荐系统[J]. 华东师范大学学报(自然科学版), 2013, 2013(3): 37-45.