RAWLSNET: Altering Bayesian Networks to Encode Rawlsian Fair Equality of Opportunity

David Liu, Zohair Shafi, Will Fleisher, Tina Eliass-Rad, Scott Alfeld

Presented at AIES'21

Full Paper

Reproducibility Instructions

The core implementation of RAWLSNET is included in the rawlsnet.py file. The parent method in this file is perturb_bayes which then calls all of the necessary helper functions.

The file college.py reproduces Figure 2 from the paper -- the core example from the paper. Future examples and experiments can be created by writing a new main function and calling perturb_bayes accordingly. Users may find the comment header for pertub_bayes helpful. For convenience the documentation is also included below:

def perturb_bayes(data):
	# Takes as input a dictionary containing : 
	# 1. Advantageous Social Position - String
	# 2. Control Variable - String
	# 3. Sensitive Variables - List of Strings
	# 4. Justified Variables - List Of Strings
	# 5. Control Input Variable - List Of Strings 
	# 6. Bayes Net Exists - "yes" / "no"
	# 7. Edge List - Continous List Of Nodes (node i and i+1 form an edge)
	# 8. CPT Values - Numpy Array Of CPT Values
	# 9. Hashed File Name - String [Optional]
	# 10. Constraint Value - Dictionary 
	# 11. Manual Constraints - list
	# 12. CPT Exists - "yes" / "no"
	# 13. File Path - Path to csv file
	# 14. Out Path - Path to write HTML plots

	# data = 
	# {

	# 	'adv_soc_pos_val' = 'Job',
	# 	'control_val' = 'College',
	# 	'sensitive_val' = ['SES'],
	# 	'justified_val' = ['Talent'],
	# 	'control_input_val' = [],
	# 	'bayes_net_exists' = 'yes',
	# 	'cpt_exists' = 'yes',
	# 	'bn_edge_list' = ["SES", "College", "SES", "Job", "Test", "College", "College", "Job"],
	# 	'cpt_values' = [
	# 					{variable: "SES", variable_card: "2", evidence: [], evidence_card: "", values: ["0.8", "0.2"]}
	# 					{variable: "Test", variable_card: "2", evidence: [], evidence_card: "", values: ["0.5", "0.5"]}
	# 					{variable: "College", variable_card: "2", evidence: ["SES", "Test"], evidence_card: "2,2", values: ["0.8", "0.4", "0.3", "0.1", "0.2", "0.6", "0.7", "0.9"]}
	# 					{variable: "Job", variable_card: "2", evidence: ["SES", "College"], evidence_card: "2,2", values: ["0.9", "0.4", "0.3", "0.1", "0.1", "0.6", "0.7", "0.9"]}
	# 				   ]
	# 	'hashed_file' = "",
	# 	'file_path' = "/somewhere/data.csv",
	# 	'out_path' = "/thatplace/results/",
	# 	'constraint_val' = "0.5",
	# 	'manual_constraints' = {'A': [[0, 0, 1, 0], [0, 0, 0, 1]], 'B': [0.1, 0.85]}
	# }

After executing the main function, all figures are exported into the /figs directory.

All of the python dependencies are specified in requirements.txt.

Implementation Notes

The implementation of RAWLSNET released here allows for categorical query variables (e.g. job) but all other variables must be binary. From a methodological perspective, categorical variables are not an issue; it simply requires additional implementation. To support categorical variables, the generate_binary_truth table function in utils.py should be replaced by a helper function that generates all permutations of the provided variables.

Contact

Please direct all questions and comments to David Liu at liu.davi@northeastern.edu