Analysis of Data Augmentation for Federated Sequential Recommendation
- 주제(키워드) Recommendation System , Federated Learning , Data Augmentation
- 주제(DDC) 006.31
- 발행기관 아주대학교 일반대학원
- 지도교수 Hyunsouk Cho
- 발행년도 2025
- 학위수여년월 2025. 2
- 학위명 석사
- 학과 및 전공 일반대학원 인공지능학과
- 실제URI http://www.dcollection.net/handler/ajou/000000034415
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
Federated learning is a trending method of solving privacy issues in recommendation systems. Due to its distributional manner of training, it brings up interesting challenges such as reducing communications, solving the recommendation quality gap between centralized training, and generalizing over users who don’t participate in the training. While data augmentations have been used in other studies to solve this problem, they have not yet been examined in the federated recommendation system. In this study, we analyze popular data augmentation methods used in sequential recommendation systems for dataset statistics and different settings for federated learning. Our results showed that both factors are crucial in determining the best augmentation strategies.
more목차
1 Introduction 1
2 Preliminary 3
2.1 Federated Learning 3
2.2 Federated Recommendation System 4
2.3 Sequential Recommendation 5
2.4 Data Augmentation 6
3 Data Augmentation for Sequential Recommendation System 7
3.1 Introduction of Data Augmentation 7
3.2 The effect of Data Augmentation on Client-wise Distribution Difference 9
4 Experiment 11
4.1 Experimental Settings 11
4.2 Statistical Heterogeneity Analysis 13
4.3 Analysis on Recommendation Accuracy 14
4.4 Analysis on Convergence Speed 15
4.5 Analysis on Generalization on Unseen Users 18
5 Related Work 21
5.1 Non-IID Federated Recommendation Systems 21
5.2 Data Augmentation for Federated Learning 22
6 Conclusion 23
References 25
