3-7-林丽-基于动态图提示的高效时空预测方法.pdf

1、基于动态图提示的高效时空预测方法主讲人：林丽目录时空预测问题时空预测问题在多场景应用下的形式01预训练时空模型解决方案预训练时空模型用于多元时空数据预测问题02动态图提示下的时空预测方法基于动态图提示的时空预测方法03实验验证验证时空预测方法的有效性、鲁棒性、效率04结论预训练时空模型的高效应用0501时空预测问题在多场景应用下的形式时空预测问题TrendPeriodicStationary1Data analysis on IoTDBHow to recognize multivariate patterns in time-series data?1 Liu Y,Zhang H,Li C,

2、et al.Timer:generative pre-trained transformers are large time series models.ICML 2024时空数据的形式Multivariate time-series data in traffic flow.Spatio-temporal data时空数据的形式02预训练时空模型解决方案预训练时空模型用于多元时空数据预测问题How to solve spatio-temporal prediction tasks?Pre-trained ModelTraffic:Taxi,BikeLogistic:DeliveryMobil

3、ity:HumanSpatio-temporal DataTuningPre-trainingDifferent tasks 1586NodeEdgeSubgraphMasked2-3Taskmismatch2 Shao Z,Zhang Z,Wang F,et al.Pre-training enhanced spatial-temporal graph neural network for multivariate time series forecasting.KDD 2022.3 Li Z,Xia L,Xu Y,et al.GPT-ST:generative pre-training o

4、f spatio-temporal graph neural networksJ.NeurIPS 2023.预训练时空预测模型Pre-trained ModelPre-trained ModelTuningPre-training,and Fine-tuningInefficiency,Knowledge ForgettingDifferent tasks Pre-training,Prompting and Fine-tuningPromptPromptHow to solve spatio-temporal prediction tasks efficiently?Pre-trained

5、Model预训练时空预测模型03动态图提示下的时空预测方法基于动态图提示的时空预测方法We propose a framework that Prompts future snapshots on Spatio-Temporal graphs for various types of downstream tasks(ProST).Spatio-Temporal modelingTask mismatchingTask matching Feature PromptSnapshot PromptOursPromptingDownstream TasksFuture SnapshotPre-tr

6、ainingPretext TasksDownstream TasksMatchingTuningPre-trained ModelPre-trained Model基于动态图提示的时空预测方法Pre-trainingPretext TasksDownstream TasksMatchingTuning1586Node-level TaskEdge-level TaskSubgraph-levelTaskSingle-node SubgraphNode-pair SubgraphSubgraphSubgraph-basedTasksMatchingPre-trained Model基于动态图提