RobinLu1209 / STGCN

Undergraduate graduation project: Spatial-Temporal Data Analysis based on Graph Convolution Network

Spatial-Temporal Data Analysis on Graph Convolutional Network

Baseline Result

Dataset	Model	MAE	Parameter
METR-LA	STGCN(pytorch)	3.982	Epoch = 1000

Data

METR-LA and PEMS-BAY

Dataset	Node num	Time	Duration	Time slot	Scene
METR-LA	207	2012.03.01~2012.6.27	4 months	5mins	Loop detecors in highway
PEMS-BAY	325	2017.01.01~2017.06.30	6 months	5mins	Sensors in Bay Area
PEMSD7	228	Workday of 2012.05-2012.06	44 days	5mins	Sensors in California

• METR-LA

Data	Info
distance_la_2012.csv	两两节点之间的距离
graph_sensor_ids	所有节点的id list
graph_sensor_locations	所有节点的坐标
metra_la.csv	每个节点在每个时刻的速度信息

Q_Traffic Dataset Link

The data provider gives 15073 central road and its neighbour information, so there are totally 45148 roads data(speed/road netwok/gps) provided. The total time slot number is 5856(61days * 24hours * 4quarter).

Filename	Dimension	Instance	Tips
traffic_speed_sub-dataset	3 * (5856*45148)	road_id = 1562548955, timeslot_id = 0, speed = 41.3480687196	No headings, sep = ' '
road_network_sub-dataset	8 * 45148(-Heading)	road_id = 1562548955, width = 30, direction = 3, snodeid = 1520445066, enodeid = 1549742690, length = 0.038, speedclass = 6, lanenum = 1	Headings, sep = '\t'
link_gps	3 * 45418	road_id = 1562548955, longtitude = 116.367557, latitude = 39.899537	No headings, sep = ' '
query_sub-dataset	61 * 6 * N	search_time = 2017-04-01 19:42:23, start_pos = (116.325461 40.036083), end_pos = (116.350811 40.090999), travel_time = 33	No headings, sep = ' ' or ','
neighbours_1km.txt	15073 * 11	road_id = xx, pre1, pre2, ..., pre5, next1, next2, ..., next5

Highways England network journey time and traffic flow data Link

Baseline Analysis

DCRNN

1. Data pre-processing
   • speed_h5py.py is used to generate speed dataset in h5 format.
   • Then, use this speed_dataset to generate train/validate/test data by code generate_training_data.py .
   • gen_adj_mx.py is used to generate road_map.
2. Train DCRNN model
   • Comand line(The version of tensorflow-gpu must be higher than tensorflow):

tmux a -t dcrnn_baidu
source activate python3.6
cd ~/workspace/GCN/DCRNN-master
python dcrnn_train.py --config_filename=data/model/dcrnn_baidu.yaml

Basic Models

1. ChebNet: Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering
2. STGCN: Spatio-Temporal Graph Convolutional Networks | For pytorch version: pytorch version
3. DCRNN: Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting
4. Multi-head Self Attention Model(AutoInt): AutoInt: Automatic Feature Interaction Learning via Self-Attentive Neural Networks CSDN reference: AutoInt：使用Multi-head Self-Attention进行自动特征学习的CTR模型

Environmental Data

1. Targeted source detection for environment data

Basic Methods

1. K-SVD in Dictionary learning There are codes and some illustration.
2. [osmnx guide]https://github.com/gboeing/osmnx-examples/tree/master/notebooks()
3. python GIS

Tips

Tensorflow and CUDA compatible combinations

version	Python version	cuDNN	CUDA
tensorflow-gpu-1.14.0	python3.6	7.6	10.0