The A Cuda-based Parallel PC Algorithm (fastPC) is implemented in PyTorch. Given a dataframe as input, fastPC discovers causal relationships between these random variables and outputs a causal graph. It can also perform multiple imputation.
Paper:
- Discovers causal relationships between random variables
- Plots causal graph
Check out the Jupyter Notebook fastPC Demo to see a demonstration of the functionality.
- Python >= 3.5
- PyTorch (tested with PyTorch 1.7.1)
- matplotlib>=3.3.3
- miceforest>=2.0.3 (https://pypi.org/project/miceforest/)
- networkx>=2.5
- numpy>=1.19.2
- pandas>=1.1.3
- scikit-learn>=0.24.1
- Optional: CUDA (tested with CUDA 11.0)
- os
- sys
- argparse
- itertools
- logging
- time
Required: data.csv (CSV file with variable names as header, column: variable, row: samples)
Optional: blacklist.txt (TXT file, each line denotes an edge that must NOT exist, separate by ",". E.g. if we know var_1 cannot cause var_2, write "var_1, var_2" in one line)
knownedges.txt (TXT file, each line denotes an edge that COULD (not must) exist that we already know direction, separate by ",". E.g. if we know var_1 - var_2, then the direction must be var_1 -> var_2, then write "var_1, var_2" in one line)
tiers.txt (TXT file, each line denotes all variables that belong to same tier, edge directions can onlt point from former to same or later tiers, later tiers can NOT point to former tiers)
The folder 'data' contains one small dataset for demonstration purposes.
Check out the Jupyter Notebook fastPC Demo to see a demonstration of the functionality.
Run fastPC.py --data yourpath/yourdata.csv to run fastPC on your own dataset(s). fastPC will discover causal relationships between random variables in the dataset, and outputs a completed partially directed acyclic graph (CPDAG).
The causal graph will be plotted as well as saved in three files:
graph_excel.csv: contains all edges in the CPDAG;
graph_excel_bidirection.csv: only the bidirection edges in the CPGAD;
graph_excel_single_direction.csv: only the single direction edges in the CPGAD.
Feel free to improve fastPC use Issues.
Corresponding Paper (peer-reviewed, open access):
Please cite this paper when using fastPC: