Code for the paper titled "Luckiness Normalized Maximum Likelihood-based Change Detection for High-dimensional Graphical Models with Missing Data".
- Datasets
All of the datasets we mentioned in the paper are in data/.
In synthetic_data/, there are complete datasets in complete/, and corresponding datasets with missing values in s01/ (10% missing values) and s025/ (25% missing values). For the datasets with missing values, files ending with _miss_1 means their missingness is at random, while files ending with _miss_2 means their missingness is in stripes.
Folder realworld_data/ includes datasets S&P 500 in sp500/ and COVID-19 in covid/. sp500/2007_2009_p200_12miss.csv and covid/2020_2021_covid.csv are preprocessed datasets we used in our experiments. The original real-world datasets are not provided here since their sizes exceed the limit.
- Scripts
Please install the Python package pyglassobind before running the scripts.
Script cpd_LNML.py outputs the value of MDL change statistics, while cpd_baseline.py outputs the gains of baseline methods. The scripts use s01_p30_*.csv as an example. All of the Python scripts can be easily run by
python *.py
- Evaluation
The folder data/synthetic_data/gain/ includes the MDL change statistics and gains of baseline methods obtained from the synthetic data.
We provide two example scripts to run the evaluation in gain/s01/p30/. Both eval_LNML.py and eval_baseline.py use benefit versus FAR as the criterion.