Awesome

MOCHE

Comprehensible Counterfactual Interpretation on Kolmogorov-Smirnov Test

Setup Experiment Environment

Use the following command to setup the experiment environment.

The command will install the dependencies required by MOCHE.

python setup.py install

Interpretation Methods

grace.py: The baseline method GRACE [1].

corner_search.py: The basline method CornerSearch [2].

mochi.py: Our proposed method, MOCHE.

baselines.py: Abstract class of interpretation methods. The baseline methods STOMP [5], Series2Graph [6], and D3 [7] are implemented in the file as well.

Run Experiments

Produce interpretations with the preference list generated by Bitmap [3]

sh bin/run_exp_bitmap.sh

Produce interpretations with the preference list generated by Spectral Residual [4]

sh bin/run_exp_spectral.sh

Produce interpretations on the failed KS test conducted on the COVID-19 dataset

python exp/exp_covid.py

Dataset

We use the dataset in the NAB repository and the COVID-19 cases in British Columbia, Canada.

Links

Bitmap Library: https://github.com/linkedin/luminol

Spectral Residual: https://docs.seldon.io/projects/alibi-detect/en/latest/api/modules.html

STOMP: https://matrixprofile.org

Series2Graph: http://helios.mi.parisdescartes.fr/~themisp/series2graph/

NAB Dataset: https://github.com/numenta/NAB/tree/master/data

COVID-19 Dataset: http://www.bccdc.ca/health-info/diseases-conditions/covid-19/data

Reference

[1] Generating Concise and Informative Contrastive Sample to Explain Neural Network Model’s Prediction

[2] Sparse and Imperceivable Adversarial Attacks

[3] Assumption-free anomaly detection in timeseries

[4] Time-Series Anomaly De-tection Service at Microsoft

[5] Matrix Profile I: All Pairs Similarity Joins for Time Series: A Unifying View that Includes Motifs, Discords and Shapelets

[6] Series2Graph: Graph-based Subsequence Anomaly Detection for Time Series

[7] Online Outlier Detection in Sensor Data Using Non-Parametric Models