The content introduces a novel generative pre-training framework, GPD, for spatio-temporal few-shot learning. It addresses challenges of data scarcity and heterogeneity in smart city applications. The framework leverages pre-training in the parameter space to facilitate knowledge transfer across different cities. By optimizing neural network parameters and utilizing a Transformer-based denoising diffusion model, GPD generates tailored neural networks guided by prompts for accurate predictions. Extensive experiments on real-world datasets demonstrate superior performance compared to state-of-the-art baselines.
다른 언어로
소스 콘텐츠 기반
arxiv.org
더 깊은 질문