ETA Prediction with Graph Neural Networks in Google Maps
- 1. ETA Prediction with Graph Neural Networks in Google Maps Austin Derrow-Pinion, Jennifer She, David Wong, Oliver Lange, Todd Hester, Luis Perez Archived 2021. 09. 10 Hyunwook Lee
- 2. Contents • Overview • Preliminaries • Methods • Experiment • Conclusion
- 3. 2 Overview: What is ETA? • ETA: Estimated time of arrival Equivalent to travel time estimation
- 4. 3 Overview
- 5. 4 Preliminaries • Goal: ETA for the supersegment S and segment s in variety of horizon T Supersegment: 𝑺 = 𝑆, 𝐸 , 𝑠 ∈ 𝑆 & 𝑒𝑖𝑗 ∈ 𝐸 • Input features speed, supersegment travel time • Real-time information is provided for 17 2-minutes windows • Historical data is provided for 5 8-minutes windows (average across the past 17 weeks) Segment-wise features – including road classifications, length, and priority • Embeddings for supersegment/segment are given
- 6. 5 Methods: What is Graph Network? • φ: update function • ρ: aggregation function • u: global attributes
- 7. 6 Methods: What is Graph Network? • Well-known GN block configurations • In this work, they utilize 3 Full GN blocks – each of them operates as encoder, processor, and decoder • Simple Spatial Attention, GCNs can be seen as one kind of (d)
- 8. 7 Methods: MetaGradients • First used in Reinforcement Learning • Set learning rate as hyperparameter update both learning rate and model parameter simultaneously • Symbols 𝜏: training examples 𝜃: model parameters η: meta-parameter (in this work, learning rate)
- 9. 8 Methods: Others • To reduce the variance • Huber Loss More robust for the outlier than MSE More sensitive/exact in range(strongly convex) • Exponential Moving Average(EMA) of Parameters 𝜃EMA = α 𝜃EMA + (1 - α) 𝜃 Can utilize more stable parameters Note: not applied directly in training
- 10. 9 Methods: Model Training • Note: We should train individual model for each h-minutes prediction • Utilize combination of various loss functions Loss functions in supersegment/segment/cumulative-segment-level 𝐿 = 𝑙𝑠𝑠 + 𝜆𝑠𝑙𝑠 + 𝜆𝑠𝑐𝑙𝑠𝑐 𝑦 𝑐,𝑗,𝑡+ℎ (𝑖) = 𝑗 𝑦𝑗,𝑡+ℎ 𝑖 𝑓𝑘 (𝑖) is free flow time
- 14. 13 Conclusion & Discussion • Simple GNN w/ other methods can achieve remarkable performance • Embedding & other costs can be reduced Meta-Learning approaches • Pre-defined supersegment can be helpful for the ETA Prediction for N routes(or segments) is much expansive than prediction for M supersegment – M << N • Where they utilize updated edge representation?
- 15. 14 Appendix • Graph Networks https://arxiv.org/pdf/1806.01261.pdf • MetaGradients https://arxiv.org/pdf/1805.09801.pdf • Author’s Explanation Video https://www.youtube.com/watch?v=a6WCZn7kOhk
Editor's Notes
- #3: 크게 내용이 많지는 않음. 어려운 내용도 없기때문에 방법론 위주의 설명을 진행할 것임 이 논문에 크게 특이한 점은 없지만, 여러가지 방법론을 통해 학습을 stabilize했음 이러한 방법론을 알아두시면 좋을 것.
- #4: 3, 45
- #5: Strong benefits in Google-Map application 기본적인 간단한 모델인 GNN에 추가적인 학습 방법론(MetaGradients, EMA)을 통한 성능 향상
- #6: Note: supersegment는 route 전체를 의미하는 것이 아니라, Typical route를 따라 연결된 도로구간입니다.
- #7: 파이는 단순 MLP로 이루어지는 경우가 많고, 로는 summation, min/max 등 여러 aggregation 방법론이 있다. Note: Graph-XX라는 식의 거의 모든 모델들은 사실 Graph-structure를 외부에서/prior로써 받아서 다룬다
- #10: Huber loss 에러가 큰 outlier에 대해서는 constant update & 에러가 작은 case에 대해서는 squared update 에러가 클 때 MSE의 단점을 커버, 에러가 작을 떄에 MSE의 장점을 가짐
- #13: 이 외에도 aggregation functio에 대한 실험, Unsupervised auxiliary loss function을 이용했을 때에 대한 실험 등이 있지만 스킵.
- #16: 3, 45