By combining parameter-efficient fine-tuning (LoRA) with multimodal fusion of urban context, you can build accurate traffic prediction models that use fewer trainable parameters without sacrificing performance.
This paper presents PEHT, a traffic prediction model that combines Transformers with urban mobility data to forecast cellular network demand. It uses LoRA to reduce parameters while a multimodal fusion strategy integrates congestion and mobility information, achieving better accuracy than existing methods on real telecom data.