文献阅读笔记 | Urban flow prediction from spatiotemporal data using machine learning: A survey

（这几天多看看survey和review好了，话说survey和review有区别么？）

原文：Xie, P., Li, T., Liu, J., Du, S., Yang, X., & Zhang, J. (2020). Urban flow prediction from spatiotemporal data using machine learning: A survey. Information Fusion, 59, 1-12.

原文链接：https://www.sciencedirect.com/science/article/pii/S1566253519303094?casa_token=h2CrTzVjno0AAAAA:LGA_QAyz6fDBc1hp5aOeMLrrqIt0hN5Ob18LQK4vZpXM_XJ-5QfYKQUMb9KEStOJ4WrdiVOt8JWx

Personal note

个人觉得这篇写的一般。

影响urban flow的因素

1. Daily flow activity patterns
2. Anomalies of flow activity patterns
3. Weather
4. Holidays

Dataset

Dataset的类型

一般都是spatiotemporal data，都具有temporal dependence和spatial correlation

1. mobile phone data
2. taxi trajectories data
3. metro/bus swiping data

Dataset的预处理

（暂略）

public urban spatiotemporal的Datasets

（作者好好哦，收集了相关的公开数据）

• NYC taxi data: This type of data can be used for citywide crowd flow prediction, taxi movement and Taxi pick-up/drop-off tasks.
• NYC bike data: Such kind of data can be used for crowd flow forecasting, traffic prediction of bike-sharing system and bike flow prediction tasks.
• San Francisco taxi data: The taxi data can be used with NYC taxi data before.
• Weather and events data: The weather and events data can be used with other traffic flow data for estimating citywide traffic volume, like [86] .
• UK traffic flow datasets: The traffic flow data is suitable for short-term traffic flow forecasting, and it has been used in [56] .
• Traffic flow data is available from the Illinois Department of Transportation: The traffic flow data can be used as UK traffic flow datasets.
• Weather and climate data: The weather and climate data are suitable for weather forecasting related tasks than urban flow prediction tasks.
• Chicago bike sharing data: The bike sharing data is often used for urban bike flow pre-
  diction.
• NSW POI data: The POI data are frequently combined with road network data, GPS trajectories for estimating citywide traffic volume.
• Road network data: The road network data is one of the vital data constituent parts for urban flow prediction.

预测的分类

• 从空间角度：citywide-level，region-level，road-level
• 从时间角度：short-term，mid-term，long-term

Techniques

statistics-based methods

• ARMA
• ARIMA
• SARIMA

尽管方法能够捕捉时间序列数据中的temporal dependence，但它们不能刻画城市流量预测问题中的spatial correlation

traditional machine learning methods

SVR，LCGBN等等。这些方法主要适用于short-time的流量估计，有很高的预测精度。但是，最近几年的爆炸性的数据增长，而传统的机器学习方法受限于挖掘大流量数据中深层的、隐含（implicit）的spatiotemporal correlations。

deep learning-based methods

Crowd flow

DeepST

Traffic flow

Public transit flow

reinforcement learning methods

一般用于traffic flow的优化问题。

transfer learning methods

一般用于数据较为短缺的case中。主要好像是找一个和case的城市类似的、数据较为充分的城市作为类比，来分析。

challenge

1. 有哪些因素影响了flow？
2. 需要不同来源的数据（交通、社会、天气数据），如何fuse这些不同类型的data作为我们的input？
3. 有时候数据会有缺失/数据较少，这也是我们需要面对的问题。