Article Details

Abstract

To address the issue of low cache hit rates and high system load caused by traditional caching strategies failing to capture the dynamic fluctuations in flight route popularity on airline travel retail platforms, this study proposes a multi-model hybrid inference service solution. The approach integrates an offline model (XGBoost) for stable prediction, while the online module performs incremental updates to adapt to real-time streaming data. The ADWIN algorithm is introduced for real-time drift detection, combined with a dynamic learning rate mechanism to accelerate model adaptation. Experiments conducted on a platform handling approximately one million daily flight search requests demonstrate that the model's daily prediction accuracy improved from 81.0% to 87.2% (a gain of 6.2 percentage points), the cache hit rate increased by 12.3 percentage points, and the system response time is reduced by 20.4%. This research provides a feasible technical pathway for cache optimization in airline platforms and offers valuable insights for real-time inference scenarios such as E-Commerce recommendation and transportation scheduling.

Keywords

multi-model hybrid; online Machine Learning; data drift; dynamic learning rate; caching strategy

References

1. [1] Liu Haolin, Zhao Ziyu, Xie Wenfei, etc Research on Machine Learning based Civil Aviation Passenger Volume Prediction Method [J]. Advances in Applied Mathematics, 2024, 13 (8): 3814-3825
2. [2] Han Jin Design and Key Technology Implementation of Distributed Cache System with Unbalanced Access Based on Redis [D]. Chengdu: University of Electronic Science and Technology of China, 2024
3. [3] KÖLKER K，ZENGERLING Z，KÜHLEN M，et al. Assessing the Impact of Contrail Avoidance Through Rescheduling on Airline Network Flows: A Case Study of North Atlantic Flights [J/OL].Transportation Research Part A，2024，187：104155（2024-07-12）.https://doi.org/10.1016/j.tra.2024.104155.
4. [4] Han Meng, Meng Fanxing, Li Chunpeng, etc A data stream drift detection method based on Kolmogorov inequality [J]. Computer Engineering and Applications, 2025, 61 (9): 102-115
5. [5] ANDREA M，MARINA L D G，ULF N，et al. Studying and Mitigating the Effects of Data Drifts on ML Model Performance at the Example of Chemical Toxicity Data [J/OL].Scientific Reports，2022，12（1）： 7244（2022-05-04）.https://link.springer.com/article/10.1038/s41598-022-09309-3.
6. [6] SREEJA A，SANGEETHA E，TANVI P. Remediating Data Drifts and Re-Establishing ML Models [J].Procedia Computer Science，2023，218：799-809.
7. [7] FERNÁNDEZ-RODRÍGUEZ J D，PALOMO E J，ORTIZ-DE-LAZCANO-LOBATO J M，et al. Dynamic Learning Rates for Continual Unsupervised Learning [J].Integrated Computer-Aided Engineering，2023，30（3）：257-273.
8. [8] ZHANG Q，WU H，TAO J，et al. Fault Diagnosis of Rolling Bearing Based on CNN With Attention Mechanism and Dynamic Learning Rate [C]//2021 International Conference on Sensing, Measurement & Data Analytics in the Era of Artificial Intelligence (ICSMD).Nanjing：IEEE，2021：1-7.
9. [9] LUO L，WANG Y，PENG H，et al. Training Restricted Boltzmann Machine With Dynamic Learning Rate [C]//2021 International Conference on Sensing，Measurement & Data Analytics in the Era of Artificial Intelligence (ICSMD).Nagoya：IEEE，2016：103-107.
10. [10] Zhao Haiyan, Rong Xiaoyu, Cao Jian, etc Overview of Concept Drift Detection Techniques for Business Processes [J]. Computer Integrated Manufacturing Systems, 2024, 30 (10): 3431-3446
11. [11] KRAUS A，VAN DER AALST W M P. Machine Learning-Based Detection of Concept Drift in Business Processes [J/OL].Process Science，2025，2（1）：5(2025-04-16).https://link.springer.com/article/10.1007/s44311-025-00012-w.
12. [12] KRZYSZTOF W，PIOTR P，TOMASZ O，et al. A New Concept Drift Detection Method Based on the Ranking of Features in a Data Stream [J].Procedia Computer Science，2023，225：3294-3302.
13. [13] Wang Jiaojiao, Ma Xiaoyu, Liu Chang, etc Predictive monitoring method for business process execution results based on XGBoost incremental implementation [J]. Computer Integrated Manufacturing Systems, 2024, 30 (8): 2756-2775
14. [14] NAVARRO J J M，DOMÍNGUEZ R M J，GARCÍA C M，et al. A Real-Time Machine Learning-Based Methodology for Short-Term Frequency Nadir Prediction in Low-Inertia Power Systems [J/OL].Neurocomputing，2025，626： 129583（2025-04-14）.https://doi.org/10.1016/j.neucom.2025.129583.
15. [15] Becoming a Premier League player Research on Theory and Application of Distributed Stream Data Analysis and Real time Machine Learning [D]. Guangzhou: Guangdong University of Technology, 2020
16. [16] PITUCKWANICH W，HORMDEE D，BOONKONG A，et al. The Implementation of a Prediction System for Sugarcane's Destruction Rate From Sugarcane Stem Borer via Hybrid Machine Learning [J].IEEE Access，2025，13：45594-45608.
17. [17] WEI L，MU W，WANG J，et al. Online Incremental Learning Algorithms for Real-Time Fault Diagnosis in Charging Stations [C]//Proceedings of the 7th International Conference on Information Technologies and Electrical Engineering. New York：ACM，2025：201-207.
18. [18] Lin Feng Research on the Application of Continuous Training in Machine Learning Platforms [D]. Shanghai: Shanghai Jiao Tong University, 2021
19. [19] JING C，SHENGYI Y，TING G，et al. Multi-Type Concept Drift Detection Under a Dual-Layer Variable Sliding Window in Frequent Pattern Mining With Cloud Computing [J/OL].Journal of Cloud Computing，2024，13（1）：25（2024-02-12）.https://link.springer.com/article/10.1186/s13677-023-00566-9.
20. [20] AGRAHARI S，SINGH K A. Comparison Based Analysis of Window Approach for Concept Drift Detection and Adaptation [J/OL].Applied Intelligence，2024，55（1）：39（2024-11-27）.https://link.springer.com/article/10.1007/s10489-024-05890-4.
21. [21] Li Zhaoen, Zhang Zhihai Research on Online Machine Learning Algorithm Based on Wasserstein Distance [J]. Chinese Science: Technical Science, 2023, 53 (7): 1031-1042
22. [22] RABIU I，SALIM N，NASSER M，et al. Drift Detection Method Using Distance Measures and Windowing Schemes for Sentiment Classification [J].Computers, Materials & Continua，2022，74（3）：6001-6017.
23. [23] MOHARRAM H，AWAD A，EL-KAFRAWY P M. Optimizing ADWIN for Steady Streams [C]//Proceedings of the 37th Annual International Conference on Machine Learning. Virtual：PMLR，2022：450-459.
24. [24] Sun Zijian, Tang Jian, Qiao Junfei Dual window concept drift detection for modeling difficult parameters in industrial processes [J]. Control Theory and Applications, 2021, 38 (12): 1979-1992
25. [25] TU J，YASIN B A，MANSOR B S N. Enhancing Efficiency and Accuracy of Optimization Techniques in Time Series Data Prediction Using Machine Learning: A Systematic Literature Review [J].Journal of Computer Science and Technology Studies，2024，6（5）：129-142.

About the authors

Tao Kai (1977-), male, Han ethnicity, from Wuhan, Hubei Province, senior engineer, master's degree, research direction: digital transformation of airlines; Yang Kun (1986-), male, Han ethnicity, from Handan, Hebei Province, bachelor's degree, research direction: big data, machine learning; Huang Zhi (1991-), male, Han ethnicity, from Dalian, Liaoning, holds a master's degree and specializes in research areas such as big models and machine learning; Wu Huan (1993-), female, Han ethnicity, born in Daqing, Heilongjiang Province, assistant engineer, bachelor's degree, research direction: civil aviation artificial intelligence; Li Yong (1985-), male, Han ethnicity, born in Jingdezhen, Jiangxi, is a senior engineer with a master's degree. His research direction is civil aviation information technology; Wang Chao (1988-), male, Han nationality, from Harbin, Heilongjiang Province, engineer, master's degree, research direction: airline e-commerce.