Missing Value Imputation for Multi-attribute Sensor Data Streams via Message Propagation

This is the repository of MPIN project which consists of supplementary material and implementation details.

Implementation Details

Requirements

• Pytorch 1.8.1
• Numpy 1.19.2
• Pandas 1.1.3
• Sklearn 0.24.1
• tsdb 0.0.7
• pypots 0.0.9
• torch_geometric

You may use " pip3 install -r requirements.txt" to install the above libraries.

Usage

Snapshot imputation for a window:

cd ./snapshot for a window;

MPIN

bash run_MPIN.sh alone 4 0.1 200 SAGE ICU 10 testMissRate 1;

baselines

bash run_FP.sh alone 4 0.5 40 ICU 10 testMissRate 1; ### FP

bash run_traditional_imputers.sh alone 4 0.5 ICU 10 testMissRate MICE 1; ### traditional imputers, e.g., MICE

bash run_neural_network_imputers.sh 4 0.5 ICU testMissRate saits 1; ### neural network based imputers, e.g., saits

Continuous imputation:

cd ./continuous;
bash run_continuous_imputation.sh data 4 0.5 200 SAGE ICU 10 testMissRate 1.0 true 0.6;

Explaination of Parameters

Snapshot imputation for a window:

e.g, bash run_MP.sh alone 4 0.5 200 SAGE ICU 10 testMissRate 1

4：window length; 0.5: missing ratio; 200: training epochs; SAGE: base model, other options such as GAT, GCN; ICU: dataset, other options such as Airquality, KDM (i.e.,Wi-Fi); 10: K value of KNN; testMissRate: effect, other options such as testWindowLen, testNumStream; 1: ratio of streams.

Continuous imputation:

bash run_continuous_imputation.sh data 4 0.5 200 SAGE ICU 10 testNumStream 1.0 true 0.6

data: update mechanism MPIN-D, other options such as state (MPIN-M), data+state (MPIN-DM), alone (MPIN-P); 4：the same meaning as above; 0.5: the same meaning as above; 200: the same meaning as above; SAGE: ...

Acknowledgements

We appreciate the work of SAITS, and their contributed codes available in here.