Reproducibility study of FairAC

An implementation of "Fair Attribute Completion on Graph with Missing Attributes" paper by D. Guo, Z. Chu, and S. Li.

To run experiments:

Install dependencies using conda

conda env create -f environment.yml

Code structure

The code is organised as follows:

• dataset/ contains the dataset files
• src/models/ contains PyTorch models for FairAC/FairGNN
• experiments/ contains notebooks to run the experiments

Run the experiments

The experiments can be run using the notebooks in the experiments/ folder. The notebooks are self-explanatory and can be run in order to reproduce the results.

Evaluation of the provided model weights

We provide the pre-trained weights from our experiments in experiments/logs. These weights can be evaluated using the experiments/run_fair_ac.ipynb notebook. The notebook is self-explanatory and can be run to evaluate the model weights.

Deepwalk embeddings

Several deepwalk embeddings are provided already. If more are needed they can be created by running:

python src/create_deepwalk_emb.py

Results

Dataset	Method	M	Acc ↑	AUC ↑	ΔSP ↓	ΔEO ↓	ΔSP+ΔEO ↓	Consistency ↑
NBA	GCN	✔	66.98 ± 1.18	76.15 ± 1.40	0.14 ± 0.13	0.57 ± 0.06	0.71 ± 0.18	2.64 ± 0.00
	ALFR	✖	64.3±1.3	71.5±0.3	2.3±0.9	3.2±1.5	5.5±2.4	-
	ALFR-e	✖	66.0±0.4	72.9±1.0	4.7±1.8	4.7±1.7	9.4±3.4	-
	Debias	✖	63.1±1.1	71.3±0.7	2.5±1.5	3.1±1.9	5.6±3.4	-
	Debias-e	✖	65.6±2.4	72.9±1.2	5.3±0.9	3.1±1.3	8.4±2.2	-
	FCGE	✖	66.0±1.5	73.6±1.5	2.9±1.0	3.0±1.2	5.9±2.2	-
	FairGNN	✔	68.39 ± 3.12	74.29 ± 1.19	2.81 ± 4.01	3.00 ± 4.07	5.81 ± 8.08	2.64 ± 0.00
	FairAC (Ours)	✔	66.51 ± 1.09	75.69 ± 1.31	0.09 ± 0.08	0.10 ± 0.00	0.19 ± 0.08	2.64 ± 0.00
Pokec-z	GCN	✔	65.10 ± 0.24	68.42 ± 0.12	1.72 ± 1.17	1.37 ± 0.51	3.08 ± 1.68	41.35 ± 0.01
	ALFR	✖	65.4±0.4	71.3±0.3	2.8±0.5	1.1±0.4	3.9±0.9	-
	ALFR-e	✖	68.0±0.6	74.0±0.7	5.8±0.4	2.8±0.8	8.6±1.2	-
	Debias	✖	65.2±0.7	71.4±0.6	1.9±0.6	1.9±0.4	3.8±1.0	-
	Debias-e	✖	67.5±0.7	74.2±0.7	4.7±1.0	3.0±1.4	7.7±2.4	-
	FCGE	✖	65.9±0.2	71.0±0.2	3.1±0.5	1.7±0.6	4.8±1.1	-
	FairGNN	✔	68.16 ± 0.59	75.67 ± 0.52	1.56 ± 0.45	3.17 ± 1.07	4.73 ± 1.47	41.35 ± 0.01
	FairAC (Ours)	✔	65.33 ± 0.30	71.20 ± 1.74	0.55 ± 0.10	0.13 ± 0.15	0.68 ± 0.09	41.33 ± 0.00
Pokec-n	GCN	✔	67.88 ± 1.46	72.86 ± 1.44	3.22 ± 1.29	5.93 ± 2.76	9.15 ± 4.05	45.93 ± 0.00
	ALFR	✖	63.1±0.6	67.7±0.5	3.05±0.5	3.9±0.6	3.95±1.1	-
	ALFR-e	✖	66.2±0.4	71.9±1.0	4.1±1.8	4.6±1.7	8.7±3.5	-
	Debias	✖	62.6±1.1	67.9±0.7	2.4±1.5	2.6±1.9	5.0±3.4	-
	Debias-e	✖	65.6±2.4	71.7±1.2	3.6±0.9	4.4±1.3	8.0±2.2	-
	FCGE	✖	64.8±1.5	69.5±1.5	4.1±1.0	5.5±1.2	9.6±2.2	-
	FairGNN	✔	67.06 ± 0.37	71.58 ± 2.58	0.55 ± 0.50	0.30 ± 0.20	0.85 ± 0.31	45.93 ± 0.00
	FairAC (Ours)	✔	67.00 ± 1.93	72.57 ± 1.68	0.11 ± 0.06	0.47 ± 0.81	0.58 ± 0.76	45.94 ± 0.02

Table 1: Comparison of FairAC with FairGNN on the nba, pokec-z and pokec-n dataset.

• The methods are applied on the GCN classifier, and the values for the baselines are taken from the original paper.
• The values for FairAC, FairGNN and GCN are taken from our experiments, that can be recreated using the notebooks under experiments/.
• The values consist of the mean and standard deviation of the metric over 3 runs on the seeds 40, 41 and 42.
• The best results are denoted in bold.