This is the code for our paper Counterfactual Representation Learning.

The repository contains three folders. The data directory stored the information about the two datasets we used in the experiments. CVAE and DCEVAE include the code for the experiments in CVAE and DCEVAE causal graphs respectively.

To obtains the results shown in table 1, you should run

python3 main.py --use_label True  --run 1

for the UF, DCEVAE, CR baselines.

For the I-DCEVAE baseline, you should run

python3 main.py --use_label True --use_real True  --run 1

For the CE baseline and our method, the comman is

python3 main.py --a_f 0.0 --u_kl 0.5 --a_h 0.1  --run 1

The results in table 2 are obtained by

python3 main.py --use_label True  --run 1
python3 main.py --use_label True --use_real True  --run 1
python3 main.py  --run 1

The results in table 3 are obtained by

python3 main.py --dataset law --a_y 1 --a_r 1 --a_d 1 --a_f 0.2 --a_h 0.4 --u_kl 1 --n_epochs 2000 --lr 1e-3 --use_label True --normalize True --run 1
python3 main.py --dataset law --a_y 1 --a_r 1 --a_d 1 --a_f 0.2 --a_h 0.4 --u_kl 1 --n_epochs 2000 --lr 1e-3 --use_label True --use_real True --normalize True --run 1
python3 main.py --dataset law --a_y 1 --a_r 1 --a_d 1 --a_f 0.2 --a_h 0.4 --u_kl 1 --n_epochs 2000 --lr 1e-3 --normalize True --run 1

The results in table 4 are obtained by

python3 main.py --dataset law --a_y 1 --a_r 1 --a_d 1 --a_f 0.15 --u_kl 1  --n_epochs 2000 --lr 1e-3 --use_label True --normalize True --run 1
python3 main.py --dataset law --a_y 1 --a_r 1 --a_d 1 --a_f 0.15 --u_kl 1  --n_epochs 2000 --lr 1e-3 --use_label True --normalize True --use_real True --run 1
python3 main.py --dataset law --a_y 1 --a_r 1 --a_d 1 --a_f 0.15 --u_kl 1 --n_epochs 2000 --lr 1e-3 --normalize True --run 1

For the experiments with path dependent counterfactual fairness, you need to add --path True and --path_attribute attr in the command. For the law school dataset, attr should be GPA or SAT. And for the UCI dataset, attr should be 0 for workingclass and 1 for education.

---

The detailed experiment results are stored in CF Data Record.xlsx. You can also find the command (including the random seed) used to get the results.

To get the Figure 5 and 9 shown in the paper, run

python3 draw.py