imputation_benchmark

The official implementation of “An Experimental Evaluation of Imputation Models for Spatial-Temporal Traffic Data”

Data and missing scenarios

We used 3 traffic datasets and 20 missing scenarios(4 missing types and 5 missing rates) to evaluate each imputation model.

$$ \begin{array} {cll} \hline \text{Dataset} & \text{Time points} & \text{Nodes} & \text{Time Interval} & \text{Time range} \\ \hline \text{PEMS04} & 16692 & 307 & 5\text{mins} &2018.1.1-2018.2.28\\ \text{PEMS08} & 17856 & 170 & 5\text{mins} &2016.7.1-2016.8.31\\ \text{Seattle} & 8760 & 323 & 1\text{hour} &2015\\ \hline \end{array} $$

$$ \begin{array} {cll} \hline \text{Missing pattern} & \text{Description}\\ \hline \text{SR-TR} & \text{Spatial random missing and Temporal random missing}\\ \text{SR-TC} & \text{Spatial random missing and Temporal continuous missing}\\ \text{SC-TR} & \text{Spatial continuous missing and Temporal random missing}\\ \text{SC-TC} & \text{Spatial continuous missing and Temporal continuous missing}\\ \hline \end{array} $$

Missing rates including 10%, 30%, 50%, 70%, 90%

---

We have summaries of existing imputation models from 2018-2024.

Constructing missing data for train and test (Optional)

Run the following command to construct the corresponding missing scenario from the raw data.

cd GenerateData 
python generator.py --config xxx.conf --missrate 0.3 --misstype SR-TR

The conf file is configured with the save path, the raw data path, and a csv file containing the node adjacencies (distances).

Two files will be generated:

1. miss_data_{misstype}_{missrate}_v2.npz
2. true_data_{misstype}_{missrate}_v2.npz

Usage

#e.g. miss_datapath = miss_data_{SR-TR}_{0.9}_v2.npz
#e.g. true_datapath = true_data_{SR-TR}_{0.9}_v2.npz
miss = np.load(miss_datapath)
mask = miss['mask'] # mask matrix (T,N,F), 0 means missing
miss_data = miss['data'] # 0 means missing (T,N,F)

true_data = np.load(true_datapath)['data'] # ground-truth (T,N,F)

You can skip the data generation stage by wrapping the data in the above form.

Train and Test imputation model

AGCRN

cd AGCRN/model
python Run.py --config ./configurations/PEMS04.conf

BRITS

cd BRITS
python input_process.py --config ./configurations/PEMS04_12_SR-TR_0.1_prepare.conf
python main.py --config ./configurations/PEMS04_12_SR-TR_0.1.conf --for_test 0

E2GAN

cd E2GAN
python train.py --config ./configurations/PEMS04_12_SR-TR_0.1.conf

GCASTN

cd GCASTN/GCASTN-main/code_data_paper_632/GCASTN
python train_GCASTN.py --config ./configurations/PEMS04.conf

IGNNK

cd IGNNK
python train.py --config ./configurations/PEMS04.conf

LATC

cd LATC
python train_LATC.py --config ./configurations/PEMS04_SR-TR_0.1.conf

mTAN

cd mTAN
python train.py --config ./configurations/PEMS04_SR-TR_0.5.conf

PriSTI

cd PriSTI/PriSTI-main
python exe_survey.py --config ./config/pems04.yaml

ASTGNN

cd ASTGNN
python train_ASTGNN.py.py --config ./configurations/PEMS04_SR-TC_70.conf

ImputeFormer

cd imputeformer
python main.py.py --config ./configurations/PEMS04.yaml --data_prefix ./miss_data --dataset PEMS04 --miss_type SR-TR --miss_rate 0.9