GenST
A spatiotemporal state prediction framework based on large language models that addresses the Forecasting Unobserved Node States (FUNS) problem through semantically guided generation. It adopts a two-stage architecture comprising a spatiotemporal VAE and a generative Transformer.
Timeline (last 90 days)
The framework proposed in the paper is named GenST, which uses the semantic descriptions of nodes themselves rather than historical data for prediction.
GenST uses a fine-tuned pretrained LLM to extract high-dimensional semantic features from node descriptions as the basis for prediction.
GenST adopts a two-stage generative architecture.
In the first stage, GenST uses a Spatio-Temporal VAE (spatio-temporal variational autoencoder) to compress spatio-temporal dynamics into a latent space.
In the second stage, GenST uses GenT (Generative Transformer) to reconstruct the future states of unobserved nodes starting from noise.
The research team evaluated GenST on six traffic datasets and two non-traffic datasets.
Experimental results show that GenST significantly outperforms existing baseline methods on zero-shot prediction tasks.
All Facts (7)
The framework proposed in the paper is named GenST, which uses the semantic descriptions of nodes themselves rather than historical data for prediction.
50%UnverifiedGenST uses a fine-tuned pretrained LLM to extract high-dimensional semantic features from node descriptions as the basis for prediction.
50%UnverifiedGenST adopts a two-stage generative architecture.
50%UnverifiedIn the first stage, GenST uses a Spatio-Temporal VAE (spatio-temporal variational autoencoder) to compress spatio-temporal dynamics into a latent space.
50%UnverifiedIn the second stage, GenST uses GenT (Generative Transformer) to reconstruct the future states of unobserved nodes starting from noise.
50%UnverifiedThe research team evaluated GenST on six traffic datasets and two non-traffic datasets.
50%UnverifiedExperimental results show that GenST significantly outperforms existing baseline methods on zero-shot prediction tasks.
50%