Efficient Mobile Traffic Prediction with Federated Learning|RISS 상세보기

국문 초록 (Abstract)

이동통신망은 새로운 융합서비스들의 요구에 맞추어 다양한 기기의 연결을 지원하며 모바일 트래픽이 폭발적으로 증가하고 있다. 이를 관리하기 위해 통신사업자들은 인공지능 기반 알고리즘을 도입하고 있으나, 중앙집중형 학습 방식은 트래픽 로그를 빈번히 중앙서버로 전송해야 하므로 네트워크 용량에 부담을 준다. 이를 개선하기 위해 연합학습과 같은 분산학습 전략의 도입을 고려하지만, 통신망에 적용하기 위해서는 다음의 주요 문제를 해결해야 한다: i) 지역별 데이터 이질성 문제로 인한 예측 성능 저하 문제 ii) 제한된 컴퓨팅 자원으로 인한 학습 연산량 문제 iii) 가중치 교환으로 인한 통신 효율성 문제. 본 논문에서는 이러한 문제들을 해결할 수 있는 두 가지 연합학습 기반의 모바일 트래픽 예측 방법을 제안한다. 첫째, 연합 분할 학습 방법을 통해 지역별 데이터 이질성 문제와 학습 연산량 문제를 해결한다. 둘째, 계층별 개인화 연합학습 방법으로 데이터 이질성과 통신 효율성 문제를 해결한다. 제안한 연합 분할 학습 방법은 중앙서버의 연산 부담을 줄이며, 클라이언트의 제한된 자원으로도 고성능 모델 학습이 가능함을 입증했다. 또한, 제안한 계층별 개인화 연합학습 방법은 각 지역의 특성을 반영하여 모델을 학습시키고 높은 예측 성능을 얻으면서, 통신 비용을 크게 절감할 수 있음을 보였다. 본 연구에서 제안한 방법들은 실제 통신사업자의 데이터셋을 사용한 성능 평가에서 예측 성능, 학습 연산량 감소, 통신 효율성 측면에서 우수한 성능을 보였다.

주요단어: 모바일 트래픽 예측, 개인화 연합 학습, 연합 분할 학습, 계층별 연합 학습, 분산 학습

번역하기

이동통신망은 새로운 융합서비스들의 요구에 맞추어 다양한 기기의 연결을 지원하며 모바일 트래픽이 폭발적으로 증가하고 있다. 이를 관리하기 위해 통신사업자들은 인공지능 기반 알고...

다국어 초록 (Multilingual Abstract)

Mobile communication networks support the connection of various devices to meet the demands of new convergence services, resulting in an explosive increase in mobile traffic. To manage such explosive traffic, telecommunications operators are adopting AI-based algorithms, primarily using centralized learning methods. However, this approach requires frequent transmission of traffic logs to a central server, which burdens network capacity. To address this, decentralized learning strategies such as federated learning have been considered. However, several key issues must be resolved for effective application in mobile networks: i) Degradation in prediction performance due to regional data heterogeneity ii) Computational load due to limited computing resources iii) Communication efficiency issues due to frequent weight exchanges. This thesis proposes two federated learning-based methods to address these issues in mobile traffic prediction. First, the Federated Split Learning method addresses regional data heterogeneity and computational load issues. Second, the Layer-wise Personalized Federated Learning method addresses data heterogeneity and communication efficiency issues. The proposed Federated Split Learning method reduces the computational burden on the central server, enabling high performance model training even with limited client resources. Additionally, the proposed Layer-wise Personalized Federated Learning method trains models by reflecting the characteristics of each region, achieving high prediction performance while significantly reducing communication costs. The proposed methods are evaluated using real datasets from a telecommunications operator, demonstrating superior performance in terms of prediction accuracy, reduced computational load, and improved communication efficiency.

Keywords: Mobile traffic prediction, personalized federated learning, layer-wise federated learning, distributed learning

번역하기

목차 (Table of Contents)

국문초록 i
ABSTRACT ii
LIST OF FIGURES v
LIST OF TABLES vi
1. INTRODUCTION 1

국문초록 i
ABSTRACT ii
LIST OF FIGURES v
LIST OF TABLES vi
1. INTRODUCTION 1
1.1. Background 1
1.2. Related Works 4
1.2.1. Mobile Traffic Prediction 4
1.2.2. Federated Learning 5
1.3. Contributions 7
1.4. Thesis Outline 8
2. Exploratory Data Analysis of Mobile Traffic 10
2.1. Call Detail Record Data 10
2.2. Mobile Traffic Data Overview 12
2.3. Mathematical Analysis 13
2.4. Visualizations and Insights 15
3. Federated Split Learning Approach 17
3.1. Problem Description 18
3.2. Methodology 20
3.2.1. Proposed Framework 20
3.2.2. Procedure of Proposed Framework 22
3.2.3. Example: Transformer Model 23
3.2.4. Geo-clustering-based Parameter Aggregation 26
3.2.5. Parameter Decoupling Through Personalized Layer 27
3.3. Results and Conclusions 28
3.3.1. Dataset and Experiment Setup 28
3.3.2. Baselines and Evaluation Metrics 29
3.3.3. Performance Comparisons 30
3.3.4. Effect of split design 32
3.3.5. Impacts of Hyper-Parameters 33
3.3.6. Conclusion 35
4. Layer-wise Personalized Federated Learning Approach 37
4.1. Key Observation 39
4.2. Problem Description 43
4.3. Methodology 45
4.3.1. Decomposition based client clustering 45
4.3.2. Layer-wise Personalized Federated Learning 47
4.3.3. Adaptive Layer-wise Freezing 52
4.4. Results and Conclusions 54
4.4.1. Federated Model and Experiment Settings 54
4.4.2. Baselines and Evaluation Metrics 55
4.4.3. Comparisons of Prediction Performance 57
4.4.4. Personalization Analysis 64
4.4.5. Communication Cost Analysis 65
4.4.6. Conclusion 66
5. Conclusions and Future Work 69
5.1. Overall Conclusions 69
5.2. Future Work 70
Bibliography 72

참고문헌 (Reference)

1. What should 6G be? In, Shuping Dang et al, Nature Electronics 3.1, pp. 20–29, , 2020

2. Federated mutual learning, Tao Shen et al, In arXiv preprint arXiv:2006.16765, , 2020

3. A speculative study on 6G In, Faisal Tariq et al, IEEE Wireless Communications 27.4, pp. 118–125, , 2020

4. Forecasting: principles and practice, Rob J Hyndman and, George Athanasopoulos, OTexts, , 2018

5. Federated optimization in heterogeneous networks, Tian Li et al, Proceedings of Machine learning and systems 2, pp. 429–450, , 2020

6. 6G ecosystem: Current status and future perspective, Jagadeesha R Bhat and, Salman A Alqahtani, IEEE Access 9, pp. 43134– 43167, , 2021

7. Deep learning in mobile and wireless networking: A survey, Chaoyun Zhang, Paul Patras and, Hamed Haddadi, In IEEE Communications surveys tutorials 21.3, pp. 2224–2287, , 2019

8. Are transformers effective for time series forecasting? In, Ailing Zeng et al, Proceedings of the AAAI conference on artificial intelligence. Vol. 37. 9. 2023, pp. 11121–11128, , 2023

9. Mobile network traffic prediction using MLP, MLPWD, and SVM, Ali Yadavar Nikravesh et al, In 2016 IEEE International Congress on Big Data (BigData Congress). IEEE. pp. 402–409, , 2016

10. Mobile traffic prediction from raw data using LSTM networks, Hoang Duy Trinh, Paolo Dini, Lorenza Giupponi and, In 2018 IEEE 29th annual international symposium on personal, indoor and mobile radio communications (PIMRC). IEEE. pp. 1827–1832, , 2018