WWW-23- Counterfactual Reasoning for Decision Model Fairness Assessment

Accepted at WWW 2023 - Poster and Demo.

Official repository of Counterfactual Reasoning for Decision Model Fairness Assessment - The Web Conference 2023 (WWW '23)

Authors: Giandomenico Cornacchia*, Vito Walter Anelli, Fedelucio Narducci, Azzurra Ragone, and Eugenio Di Sciascio;

Setup Instructions

Following, the instruction to install the correct packages for running the experiments. run:

#(Python version used during experiments 3.8.10)
python3 -m venv venv
source venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt

Training, test models and generate CF samples

To train, evaluate models, and generate counterfactual samples for each dataset and Classifier, run:

#sudo chmod +x run_all_generation.sh + User_password if Permission Denied 
./run_all_generation.sh

or chose one of the config file for a specific dataset, and run:

python3 -u mainGenerate.py {config/{dataset}/...}.yml

Result can be found in the folder Results/{dataset}/{model}/{SF}/Genetic.pickle

N.B.: Experiments can take several days for each dataset and model. To speed up inference in the Model_Y and Counterfactual Generator loop for some scikit-learn models (e.g. SVM), the "skearnex" library from intel was used. The metrics' results may vary slightly from the scikit-learn models.

Counterfactual metrics evaluation

To evaluate the proposed metric for each dataset, simply run the following command:

python3 -u mainEvaluate.py

The script will display Counterfactual Fairness metrics in Table 3.

Dataset	Target Classifier	AUC	ACC	DSP	DEO	DAO	CFlips@10 Privileged	CFlips@50 Privileged	CFlips@100 Privileged	CFlips@10 Unprivileged	CFlips@50 Unprivileged	CFlips@100 Unprivileged	$\Delta$ CFlips@10	$\Delta$ CFlips@50	$\Delta$ CFlips@100	nDCCF@10 Privileged	nDCCF@50 Privileged	nDCCF@100 Privileged	nDCCF@10 Unprivileged	nDCCF@50 Unprivileged	nDCCF@100 Unprivileged	$\Delta$ nDCCF@10	$\Delta$ nDCCF@50	$\Delta$ nDCCF@100
ADULT	LR	0.9078	0.8099	0.2947	0.0546	0.1241	12.332	10.886	10.212	66.353	72.932	77.165	54.021	62.046	66.953	0.8678	0.8849	0.886	0.3522	0.2913	0.2497	0.5156	0.5936	0.6363
ADULT	SVM	0.9073	0.8541	0.1769	0.0644	0.0692	6.752	7.533	7.742	77.095	80.973	81.372	70.343	73.44	73.63	0.9306	0.9258	0.9171	0.2474	0.2042	0.1948	0.6832	0.7216	0.7223
ADULT	LGB	0.9304	0.8658	0.1850	0.0379	0.0569	9.195	8.541	8.781	65.918	76.605	79.697	56.723	68.064	70.916	0.9049	0.9124	0.9049	0.3611	0.2633	0.2272	0.5438	0.6491	0.6777
ADULT	XGB	0.9314	0.8698	0.1884	0.0635	0.0680	10.011	8.788	9.07	64.796	76.243	79.512	54.785	67.455	70.442	0.8968	0.9088	0.9014	0.3708	0.2677	0.2298	0.526	0.6411	0.6716
ADULT	AdvDeb	0.9123	0.8512	0.1151	0.1399	0.0879	30.046	34.488	34.968	36.11	38.694	43.041	6.064	4.206	8.073	0.7016	0.6668	0.6537	0.6427	0.6199	0.5812	0.0589	0.0469	0.0725
ADULT	lferm	0.9031	0.8428	0.1448	0.0194	0.0386	31.459	28.632	24.965	31.764	47.464	57.47	0.305	18.832	32.505	0.6857	0.7062	0.7314	0.6864	0.5632	0.4701	0.0007	0.143	0.2613
ADULT-DEB	LR	0.8233	0.7367	0.1567	0.0695	0.0693	8.438	10.838	13.192	54.816	57.521	57.047	46.378	46.683	43.855	0.9239	0.9012	0.8736	0.464	0.4332	0.4303	0.4599	0.468	0.4433
ADULT-DEB	SVM	0.8189	0.7395	0.1062	0.0140	0.0152	11.937	16.377	17.379	31.305	33.869	35.385	19.368	17.492	18.006	0.8871	0.8468	0.8295	0.6661	0.6616	0.6449	0.221	0.1852	0.1846
ADULT-DEB	LGB	0.8596	0.8371	0.1093	0.0470	0.0356	4.624	9.419	12.848	66.966	74.223	73.445	62.342	64.804	60.597	0.9578	0.9182	0.8815	0.3720	0.2863	0.2794	0.5858	0.6319	0.6021
ADULT-DEB	XGB	0.8578	0.8375	0.1056	0.0400	0.0304	1.803	3.152	6.523	81.289	88.9	84.48	79.486	85.748	77.957	0.9804	0.9711	0.9386	0.2183	0.1378	0.1599	0.7621	0.8333	0.7787
ADULT-DEB	AdvDeb	0.8309	0.8195	0.0957	0.0326	0.0282	17.041	20.686	23.588	44.315	52.371	56.786	27.274	31.685	33.198	0.8425	0.8055	0.7735	0.5852	0.5031	0.4566	0.2573	0.3024	0.3169
ADULT-DEB	lferm	0.8017	0.7953	0.0639	0.0179	0.0186	8.943	13.316	16.561	47.036	54.87	55.83	38.093	41.554	39.269	0.9248	0.8809	0.8452	0.5618	0.4791	0.4584	0.363	0.4018	0.3868
CRIME	LR	0.9248	0.8700	0.6535	0.3294	0.3438	2.857	3.429	3.714	75.286	81.914	85.043	72.429	78.485	81.329	0.9688	0.9656	0.9564	0.2659	0.2015	0.1688	0.7029	0.7641	0.7876
CRIME	SVM	0.9288	0.8700	0.6395	0.3843	0.3390	6.667	5.917	5.671	73.38	80.676	84.437	66.713	74.759	78.766	0.9334	0.939	0.9349	0.2858	0.2157	0.1781	0.6476	0.7233	0.7568
CRIME	LGB	0.9168	0.8400	0.6363	0.2824	0.3525	5.455	5.818	5.636	74.571	80.229	83.693	69.116	74.411	78.057	0.9432	0.9417	0.9364	0.2875	0.2207	0.1842	0.6557	0.721	0.7522
CRIME	XGB	0.9099	0.8500	0.6568	0.2941	0.3656	4.762	5.429	5	73.38	80.113	83.712	68.618	74.684	78.712	0.9505	0.9469	0.943	0.2938	0.2216	0.1844	0.6567	0.7253	0.7586
CRIME	AdvDeb	0.9008	0.8100	0.5501	0.1882	0.2732	7.5	6.875	6.969	69	77.743	80.857	61.5	70.868	73.888	0.9302	0.931	0.9237	0.3396	0.2506	0.2146	0.5906	0.6804	0.7091
CRIME	lferm	0.9100	0.8400	0.6125	0.2941	0.3278	3.182	6	6.636	64.412	71.647	75.147	61.23	65.647	68.511	0.9679	0.9439	0.9306	0.3695	0.3045	0.2681	0.5984	0.6394	0.6625
GERMAN	LR	0.8186	0.7600	0.1187	0.1400	0.1657	4.4	6.56	7.402	30.000	28.000	34.000	25.6	21.44	26.598	0.9656	0.9417	0.9267	0.7289	0.7197	0.6728	0.2367	0.2220	0.2539
GERMAN	SVM	0.8109	0.7600	0.0449	0.0300	0.0892	6.667	9.167	10	$0^*$	$0^*$	$0^*$	$\mathbf{6.667^\dagger}$	9.167	10	0.9276	0.9124	0.8985	$0^*$	$0^*$	$0^*$	0.9276	0.9124	0.8985
GERMAN	LGB	0.7614	0.7500	0.1632	0.1900	0.1117	5.385	7.231	9.538	30.000	44.000	50.50	24.615	36.769	40.962	0.9460	0.9329	0.9071	0.7193	0.5990	0.5313	0.2267	0.3339	0.3758
GERMAN	XGB	0.7871	0.7900	0.0518	0.0400	0.0296	4.545	6.545	8	23.333	24.000	26.333	18.788	17.455	18.33	0.9616	0.9414	0.9206	0.7448	0.7569	0.7365	0.2168	0.1845	0.1841
GERMAN	AdvDeb	0.7371	0.7300	0.1809	0.2200	0.1267	4.167	4.333	7.167	53.333	59.333	59.000	49.166	55	51.833	0.9661	0.9608	0.9309	0.4389	0.4103	0.4102	0.5272	0.5505	0.5207
GERMAN	lferm	0.7605	0.7200	0.0355	0.0200	0.0746	10	3	7.5	$0^*$	$0^*$	$0^*$	10	3	7.5	0.8552	0.9402	0.91	$0^*$	$0^*$	$0^*$	0.8552	0.9402	0.91

Generate toy figures

To generate Figure 1, run the following command:

python3 -u T-SNE.py

Image are saved in folder ./figures/3D TSNE {model} {unpriv/priv} f(x)=0.svg

N.B.: Images can be slightly different, based on seed, python and libraries versions. Neverthless, they are only toys image that will return the same conceptual findings.

License

This repo uses APACHE2 License.

To cite our work:

@inproceedings{10.1145/3543873.3587354,
author = {Cornacchia, Giandomenico and Anelli, Vito Walter and Narducci, Fedelucio and Ragone, Azzurra and Di Sciascio, Eugenio},
title = {Counterfactual Reasoning for Decision Model Fairness Assessment},
year = {2023},
isbn = {9781450394192},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3543873.3587354},
doi = {10.1145/3543873.3587354},
abstract = {The increasing application of Artificial Intelligence and Machine Learning models poses potential risks of unfair behaviour and, in the light of recent regulations, has attracted the attention of the research community. Several researchers focused on seeking new fairness definitions or developing approaches to identify biased predictions. These approaches focus solely on a discrete and limited space; only a few analyze the minimum variations required in the user characteristics to ensure a positive outcome for the individuals (counterfactuals). In that direction, the methodology proposed in this paper aims to unveil unfair model behaviors using counterfactual reasoning in the case of fairness under unawareness. The method also proposes two new metrics that analyse the (estimated) sensitive information of counterfactual samples with the help of an external oracle. Experimental results on three data sets show the effectiveness of our approach for disclosing unfair behaviour of state-of-the-art Machine Learning and debiasing models. Source code is available at https://github.com/giandos200/WWW-23-Counterfactual-Fair-Opportunity-Poster-.},
booktitle = {Companion Proceedings of the ACM Web Conference 2023},
pages = {229–233},
numpages = {5},
keywords = {Audit, Bias, Counterfactual Reasoning, Fairness},
location = {Austin, TX, USA},
series = {WWW '23 Companion}
}