This list can be considered outdated. For a more up-to-date list, check: https://github.com/lixus7/Time-Series-Works-Conferences

Traffic Prediction

Traffic prediction is the task of predicting future traffic measurements (e.g. volume, speed, etc.) in a road network (graph), using historical data (timeseries).

Things are usually better defined through exclusions, so here are similar things that I do not include:

NYC taxi and bike (and other similar datsets, like uber), are not included, because they tend to be represented as a grid, not a graph.
Predicting human mobility, either indoors, or through checking-in in Point of Interest (POI), or through a transport network.
Predicting trajectory.
Predicting the movement of individual cars through sensors for the purpose of self-driving car.
Traffic data imputations.
Traffic anomaly detections.

The papers are haphazardly selected.

Summary

A tabular summary of paper and publically available datasets. The paper is reverse chronologically sorted. NO GUARANTEE is made that this table is complete or accurate (please raise an issue if you spot any error).

paper	venue	published date	# other datsets	METR-LA	PeMS-BAY	PeMS-D7(M)	PeMS-D7(L)	PeMS-04	PeMS-08	LOOP	SZ-taxi	Los-loop	PeMS-03	PeMS-07	PeMS-I-405	PeMS-04(S)	TOTAL open
		TOTAL		38	28	6	3	3	3	3	2	2	1	1	1	1	95
G-SWaN	IoTDI	9 May 23		1	1			1	1								4
SCPT	ArXiv	9 May 23	1	1	1	1											4
MP-WaveNet	ArXiv	9 May 23	1	1													2
GTS	ICLR	4 May 21	1	1	1												2
FASTGNN	TII	29 Jan 21				1											1
HetGAT	JAIHC	23 Jan 21		1	1												2
GST-GAT	IEEE Access	6 Jan 21		1	1												2
CLGRN	arXiv	4 Jan 21	3	1													1
DKFN	SIGSPATIAL	3 Nov 20		1						1							2
STGAM	CISP-BMEI	17 Oct 20		1	1												2
ARNN	Nat. Commun	11 Sept 20		1													1
ST-TrafficNet	ELECGJ	9 Sept 20		1	1												2
M2	J. AdHoc	1 Sept 20		1	1												2
H-STGCN	KDD	23 Aug 20															0
SGMN	J. TRC	20 Aug 20		1	1												2
GDRNN	NTU	16 Aug 20		1											1		2
ISTD-GCN	arXiv	10 Aug 20		1	1												2
GTS	UCONN	3 Aug 20		1	1												2
FC-GAGA	arXiv	30 Jul 20		1	1												2
STGAT	IEEE Access	22 Jul 20		1	1												2
STNN	T-ITS	16 Jul 20															0
AGCRN	arXiv	6 Jul 20						1	1								2
GWNN-LSTM	J. Phys. Conf. Ser.	20 Jun 20		1													1
A3T-GCN	arXiv	20 Jun 20									1	1					2
TSE-SC	Trans-GIS	1 Jun 20		1	1												2
MTGNN	arXiv	24 May 20		1	1												2
ST-MetaNet+	TKDE	19 May 20		1	1												2
STGNN	WWW	20 Apr 20		1	1												2
STSeq2Seq	arXiv	6 Apr 20		1	1												2
DSTGNN	arXiv	12 Mar 20		1													1
RSTAG	IoT-J	19 Feb 20		1	1												2
GMAN	AAAI	7 Feb 20			1												1
MRA-BGCN	AAAI	7 Feb 20		1	1												2
STSGCN	AAAI	7 Feb 20						1	1				1	1			4
SLCNN	AAAI	7 Feb 20		1	1	1											3
DDP-GCN	arXiv	7 Feb 20															0
R-SSM	ICLR	13 Jan 20		1													1
GWNV2	arXiv	11 Dec 19		1	1												2
DeepGLO	NeurIPS	8 Dec 19	1			1											1
STGRAT	arXiv	29 Nov 19		1	1												2
TGC-LSTM	T-ITS	28 Nov 19								1							1
DCRNN-RIL	TrustCom/BigDataSE	31 Oct 19		1	1												2
L-VGAE	arXiv	18 Oct 19		1													1
T-GCN	T-ITS	22 Aug 19									1	1					2
GWN	IJCAI	10 Aug 19		1	1												2
ST-MetaNet	KDD	25 Jul 19		1													1
MRes-RGNN-G	AAAI	17 Jul 19		1	1												2
CDSA	arXiv	23 May 19		1													1
STDGI	ICLR	12 Apr 19		1													1
ST-UNet	arXiv	13 Mar 19		1		1	1										3
3D-TGCN	arXiv	3 Mar 19			1	1	1										3
ASTGCN	AAAI	27 Jan 19						1	1								2
PSN	T-ITS	17 Aug 18														1	0
GaAN	UAI	6 Aug 18	2	1													1
Seq2Seq Hybrid	KDD	19 Jul 18															0
STGCN	IJCAI	13 Jul 18				1	1										2
DCRNN	ICLR	30 Apr 18		1	1												2
SBU-LSTM	UrbComp	14 Aug 17								1							1
GRU	YAC	5 Jan 17			1												1

Performance

METR-LA MAE@60 mins

PeMS-BAY MAE@60 mins

NOTES: The experimental setttings may vary. But the common setting is:

Observation window = 12 timesteps
Prediction horizon = 1 timesteps
Prediction window = 12 timesteps
Metrics = MAE, RMSE, MAPE
Train, validation, and test splits = 7/1/2 OR 6/2/2

However, there are many caveats:

Some use different models for different prediction horizon.
Some use different batch size when testing previous models, as they increase the observation and prediction windows from previous studies, and have difficulties fitting it on GPU using the same batch size.
Regarding adjacency matrix, some derive it using Gaussian RBF from the coordinates, some use the actual connectivity, some simply learn it, and some use combinations.
Some might also add more context, such as time of day, or day of the week, or weather.
DeepGLO in particular, since it is treating it as a multi-channel timeseries without the spatial information, use rolling validation,
Many different treatment of missing datasets, from exclusion to imputations.

Dataset

Publically available datasets and where to find them.

METR-LA DCRNN Google Drive; DCRNN Baidu; Sensor coordinates and adjacency matrix, also from DCRNN
California department of transportation (Caltrans) Performance Measurement System (PeMS). The website is: http://pems.dot.ca.gov/. From the website: The traffic data displayed on the map is collected in real-time from over 39,000 individual detectors. These sensors span the freeway system across all major metropolitan areas of the State of California
- PeMS-BAY DCRNN Google Drive; DCRNN Baidu
Sensor coordinates and adjacency matrix, DCRNN github
- PeMS-D7(M) PKUAI26 STGCN Github
- PeMS-D7(L)
- PeMS-04 [ATSGCN