This repository contains code used to audit the stability of personality predictions made by two algorithmic hiring systems

Last update: Oct 27, 2022

Related tags

Deep Learning hiring-stability-audit

Overview

Stability Audit

This repository contains code used to audit the stability of personality predictions made by two algorithmic hiring systems, Humantic AI and Crystal. This codebase supports the 2021 manuscript entitled "External Stability Auditing to Test the Validity of Personality Prediction in AI Hiring," authored by Alene K. Rhea, Kelsey Markey, Lauren D'Arinzo, Hilke Schellmann, Mona Sloane, Paul Squires, and Julia Stoyanovich.

Code

The Jupyter notebook analysis.ipynb reads in the survey and system output data, and performs all stability analysis. The notebook begins with a demographic summarization, and then estimates stability metrics for each facet experiment as described in the manuscript.

Spearman's rank correlation is used to measure rank-order stability, two-tailed Wilcoxon signed rank testing is used to measure locational stability, and normalized L1 distance is used to measure total change across each facet. Medians of each facet treatment are estimated as well. Results are saved to the results directory, organized by metric and by system (Humantic AI and Crystal). Subgroup analysis is performed for rank-order stability and total change. Highlighting is employed to indicate correlations below 0.95 and 0.90, and Wilcoxon p-values below the Bonferroni and Benjamini-Hochberg corrected thresholds. Scatterplots are produced to compare the outputs from each pair of facet treatments. Boxplots illustrate total change. Boxplots comparing relevant subgroup analysis for each facet are produced as well.

Data

Survey

Anonymized survey results are saved in data/survey.csv. Columns described in the table below.

Column	Type	Description	Values
Participant_ID	str	Unique ID used to identify participant.	"ID2" - "ID101" (missing IDs indicate potential subjects were screened out of participation)
gender	str	Participant gender, as reported in the survey. Pre-processed to mask rare responses in order to preserve anonymity.	["Male" "Female" "Other Gender"]
race	str	Participant race, as reported in the survey. Pre-processed to mask rare responses in order to preserve anonymity. Empty entries indicates participants declined to self-identify their race in the survey.	["Asian" "White" "Other Race" NaN]
birth_country	str	Participant birth country, as reported in the survey. Pre-processed to mask rare responses in order to preserve anonymity. Empty entries indicates participants declined to provide their birth country in the survey.	["China" "India" "USA" "Other Country" NaN]
primary_language	str	Primary language of participant, as reported in the survey.	["English" "Other Langauge"]
resume	bool	Boolean flag indicating whether participant provided a resume in the survey.	["True" "False"]
linkedin	bool	Boolean flag indicating whether participant provided a LinkedIn in the survey.	["True" "False"]
twitter	bool	Boolean flag indicating whether participant provided a public Twitter handle in the survey.	["True" "False"]
linkedin_in_orig_resume	bool	Boolean flag indicating whether participant included a reference to their LinkedIn in the resume they submitted. Empty entries indicate participants did not submit a resume.	["True" "False" NaN]
orig_embed_type	str	Description of the method by which the participant referenced their LinkedIn in their submitted resume. Empty entries indicate participant did not submit a resume containing a reference to LinkedIn.	["Full url hyperlinked" "Full url not hyperlinked" "Text hyperlinked" "Other not hyperlinked" NaN]
orig_file_type	str	Filetype of the resume submitted by the participant. Empty entries indicate participants did not submit a resume.	["pdf" "docx" "txt" NaN]

Humantic AI and Crystal Output

Output from Humantic AI and Crystal is saved in the data directory. Each run is saved as a CSV and is named with its Run ID. Tables 3 and 4 in the manuscript (reproduced below) provide details of each run. Each file contains one row for each submitted input. Participant_ID provides a unique key, and output_success is a Boolean flag indicating that the system successfully produced output from the given input. Wherever output_success is true, there will be numeric predictions for each trait. Crystal results contain predictions for DiSC traits, and Humantic AI results contain predictions for DiSC traits and Big Five traits.

Run ID	System	Description	Run Dates
HRo1	Humantic AI	Original Resume	11/23/2020 - 01/14/2021
HRi1	Humantic AI	De-Identified Resume	03/20/2021 - 03/28/2021
HRi2	Humantic AI	De-Identified Resume	04/20/2021 - 04/28/2021
HRi3	Humantic AI	De-Identified Resume	04/20/2021 - 04/28/2021
HRd1	Humantic AI	DOCX Resume	03/20/2021 - 03/28/2021
HRu1	Humantic AI	URL-Embedded Resume	04/09/2021 - 04/11/2021
HL1	Humantic AI	LinkedIn	11/23/2020 - 01/14/2021
HL2	Humantic AI	LinkedIn	08/10/2021 - 08/11/2021
HT1	Humantic AI	Twitter	11/23/2020 - 01/14/2021
HT2	Humantic AI	Twitter	08/10/2021 - 08/11/2021
CRr1	Crystal	Raw Text Resume	03/31/2021 - 04/02/2021
CRr2	Crystal	Raw Text Resume	05/01/2021 - 05/03/2021
CRr3	Crystal	Raw Text Resume	05/01/2021 - 05/03/2021
CRp1	Crystal	PDF Resume	11/23/2020 - 01/14/2021
CL1	Crystal	LinkedIn	11/23/2020 - 01/14/2021
CL2	Crystal	LinkedIn	09/13/2020 - 09/16/2021

This repository contains code used to audit the stability of personality predictions made by two algorithmic hiring systems

Related tags

Overview

Stability Audit

Code

Data

Survey

Humantic AI and Crystal Output

Owner

Data, Responsibly

Source code of all the projects of Udacity Self-Driving Car Engineer Nanodegree.

PyMatting: A Python Library for Alpha Matting

COVID-VIT: Classification of Covid-19 from CT chest images based on vision transformer models

Latent Network Models to Account for Noisy, Multiply-Reported Social Network Data

Tensorflow port of a full NetVLAD network

Paddle-Skeleton-Based-Action-Recognition - DecoupleGCN-DropGraph, ASGCN, AGCN, STGCN

Official implementation of Deep Convolutional Dictionary Learning for Image Denoising.

A simple, fully convolutional model for real-time instance segmentation.

Refactoring dalle-pytorch and taming-transformers for TPU VM

HODEmu, is both an executable and a python library that is based on Ragagnin 2021 in prep.

Learnable Multi-level Frequency Decomposition and Hierarchical Attention Mechanism for Generalized Face Presentation Attack Detection

Some toy examples of score matching algorithms written in PyTorch

Code for our paper "Multi-scale Guided Attention for Medical Image Segmentation"

A simple Neural Network that predicts the label for a series of handwritten digits

Novel Instances Mining with Pseudo-Margin Evaluation for Few-Shot Object Detection

Accelerating BERT Inference for Sequence Labeling via Early-Exit

:fire: 2D and 3D Face alignment library build using pytorch

PlenOctrees: NeRF-SH Training & Conversion

A small demonstration of using WebDataset with ImageNet and PyTorch Lightning

Spline is a tool that is capable of running locally as well as part of well known pipelines like Jenkins (Jenkinsfile), Travis CI (.travis.yml) or similar ones.