A Dual-Architecture Framework for Context-Aware Video Recommendations with Dynamic Resource Optimization

This paper introduces a dual-architecture framework that optimizes video recommendations through dynamic switching between MobileNetV2 and YOLOv8 models. Evaluation on 2.3 million user interactions demonstrates that MobileNetV2 achieves 94.5% model size reduction (14.2MB vs 257.8MB) and 56.3% lower GPU memory usage (2.1GB vs 4.8GB) while maintaining 87.3% accuracy in mobile environments. The YOLOv8 component excels in complex scenarios with 89.2% accuracy, notably in crowded scenes (+8.9%) and low-light conditions (+5.3%). Our dynamic switching mechanism improves system throughput by 21.4% (3,400 vs 2,800 users/second) and reduces cold-start latency by 18.8%. The framework demonstrates robust performance across content types, achieving 89.2% accuracy for short-form videos and 84.5% for live streams, while maintaining performance stability with minimal variance. These results establish new benchmarks for resource-efficient video recommendation systems, proving that adaptive architecture switching can effectively balance recommendation accuracy with computational efficiency.