A Hybrid Deep Learning Architecture for Content Request Prediction in the Internet of Vehicles

Area of Science:

• Computer Science
• Artificial Intelligence
• Network Engineering

Background:

• Low-latency content delivery is critical for advanced Internet of Vehicles (IoV) applications like autonomous driving.
• Existing prediction and caching methods struggle with dynamic vehicular mobility and demand, failing to capture both short-term and long-term patterns.

Purpose of the Study:

• To develop a novel hybrid deep learning architecture for IoV content delivery.
• To jointly model fine-grained temporal dynamics and global correlations in vehicular content requests.
• To optimize proactive and adaptive cache placement at roadside units (RSUs).

Main Methods:

• Integration of Long Short-Term Memory (LSTM) networks and Transformer encoders for content request prediction.
• Application of a reinforcement learning (RL)-based caching policy utilizing the hybrid model's predictions.
• End-to-end optimization framework for cache placement in roadside units.

Main Results:

• The hybrid LSTM-Transformer model significantly improves cache hit ratio and reduces retrieval latency.
• Demonstrates superior prediction accuracy compared to LSTM-only, Transformer-only, LFU, and LRU baselines.
• Ablation studies confirm the complementary benefits of hybrid components, enhancing convergence and robustness.

Conclusions:

• The proposed hybrid deep learning approach effectively addresses the limitations of existing methods in dynamic IoV environments.
• This strategy enables more efficient and reliable content delivery for critical vehicular services.
• The model offers improved performance and robustness across diverse demand scenarios in the IoV.