Data-Driven Optimization of Healthcare Recommender System Retraining Pipelines in MLOps with Wearable IoT Data

Area of Science:

• Artificial Intelligence
• Machine Learning Operations (MLOps)
• Edge Computing

Background:

• Personalized healthcare recommender systems are deployed on edge AI devices via wearables.
• Cloud servers train base models, optimized for reduced data on edge devices.
• Model drift, a decline in accuracy over time, impacts user experience due to outdated predictions.

Purpose of the Study:

• To address model drift in personalized healthcare recommender systems.
• To propose a dynamic data management strategy for adaptive model updates.
• To enhance resource efficiency and maintain model accuracy during retraining.

Main Methods:

• Implemented a dynamic data management strategy within an automated machine learning operations (MLOps) pipeline.
• Utilized data reduction and feature selection algorithms to preserve base model performance.
• Validated the approach using FitRec wearable data for personalized fitness recommendations.

Main Results:

• Achieved improved computational efficiency during model retraining.
• Sustained model accuracy despite incorporating new data and addressing drift.
• Demonstrated effective mitigation of data drift and enhanced resource efficiency.

Conclusions:

• The dynamic data management strategy ensures faster training and sustained performance of base models for edge AI.
• This approach provides a robust solution for continuously refining personalized recommendation services.
• Adaptive model updates are crucial for aligning recommendations with evolving user preferences in IoT environments.