MetaTP: Traffic Prediction with Unevenly-Distributed Road Sensing Data via Fast Adaptation
summary
MetaTP claims it predicts traffic from sparse, uneven smartphone GPS by combining a two-layer interpolation that maps irregular samples to uniform time points, an LSTM for sequence modeling, and temporal and spatial memory networks that store global patterns. A meta-learning outer loop learns initial weights from source regions so models fast adapt with only a few examples in target regions. Results report improvements on two real-world datasets (Stampede, PeMS) versus standard pretrain–fine-tune.
AI-written summary · about
chapters (10)
- 0:00Intro
- 0:20Motivations & Challenges
- 0:58System Overview
- 1:30Methodology: TP-Net
- 2:07Methodology: Interpolation
- 2:56Methodology: Memory Networks
- 3:54Knowledge Transfer
- 4:43Proposed Meta-knowledge Learner
- 5:19Experiment: Stampede Dataset
- 5:35Experiment: Introduction of PeMS Dataset
related talks
07mDecentralized Attention-based Personalized Human Mobility Prediction
06mPMF: A Privacy-preserving Human Mobility Prediction Framework via Federated Learning
32mData&AI in TomTom: How are are transforming mobility
31mML time series forecasting methods
46mML Use Cases at Think Big Analytics
28mEstimating Traffic Flow to find New Drive Thru Locations
38mSampling and Modelling the World in Real-Time for Mobility
31mDeep learning for self-driving vehicle technology