Reducing Hiring Bias with AI
In collaboration with:
Overview
This project, “Enhancing Hiring Speed and Diversity with AI-driven Solutions” sponsored by Included focuses on addressing the issues of existing applicant screening processes for recruiting. Our research has found that many organizations still rely on manual resume screenings which not only have high manual labor costs, but also introduces implicit human biases to the applicant screening processes. For organizations that use ATS (applicant tracking systems) that employ AI to screen applicants, the AI may be affected by biased historical data. Our project solution remedies these issues two-fold: first, the LLM algorithm does keyword matching only on an applicant’s qualifications and second, the solution allows human recruiters to choose from a ranked list of qualified applicants and be made aware of their reviewing habits which may be biased.
Problem Statement
In today's professional world, hiring is a pivotal factor for organizational growth, regardless of the size of the company. More and more businesses are realizing the business value of diverse talent and want to include it in their teams. However, there's a common challenge that complicates this effort. Recruiters and hiring managers need to provide a diverse pool of candidates, but they often lack access to demographic data to avoid bias or discrimination.
The opportunity here is that AI-driven solutions can enable hiring managers and recruiters to have a diverse pool of qualified candidates.
Methodologies
We aimed to develop a candidate screening platform that ranks candidates based on merit and highlights potential biases in the hiring process. We mainly focused our approach towards understanding the experience of recruiters in the screening process and using what we learned to build a platform that helps them be more unbiased while meeting their candidate screening expectations.
We conducted interviews with multiple recruiters and identified features that they consider important in the screening process. Additionally, we used resume review sessions to gain insights on how recruiters look at resume. Using the findings from these methods, we leveraged AI-techniques to extract relevant identified features from job descriptions and resumes, enabling an automated and objective evaluation of candidate suitability.
The primary motivation behind this is to mitigate the inherent biases that often arise in traditional screening processes, where subjective judgments and preconceived notions can lead to unfair candidate evaluations. By using AI-powered entity extraction and matching algorithms, our platform aims to provide a more objective and data-driven assessment of candidates, thereby promoting fairness and meritocracy in the hiring process.
Benefits
-
Solution built on research
-
Objective candidate evaluation based on relevant qualifications and job requirements
-
Reduced potential for unconscious biases in the screening process
-
Transparent scoring system that provides insight into the evaluation criteria
-
Ability to adjust feature weightage and priorities based on specific hiring needs
-
Tracking of user interactions to identify potential biases and provide feedback
Research Methods
Interviews
For the semi-structured interviews, we initially chose this method because our problem space required in-depth answers to the challenges that recruiters go through while they are screening resumes which we felt wouldn’t be captured by methods like surveys. Our objectives were to understand the experience of recruiters as they screen through resumes including, the processes they use, the challenges that they face as they screen through applicants, the strategies they use to maintain a fair and unbiased process, and how they view the role of technology in their screening process. To address these objectives, we conducted interviews with 6 recruiters that made up three categories, international recruiters, agency recruiters, and company recruiters. Each interview took about 30-50 minutes and was done online or in-person depending on the interviewee’s preference.
Resume Review
After doing our interviews, we quickly realized that we needed more information regarding the explicit criteria that recruiters call out on resumes when they are screening through them. To accomplish this, we conducted resume review sessions with 3 recruiters with 4 sample resumes followed by a debrief to ask to follow up questions. For this method our objective was to know the specific areas that recruiters look for when they screen through a resume. To meet this objective, recruiters were asked to take the sample resumes and review them independently, while highlighting categories that stood out to them which took about 30 minutes on average.
Research Findings
01.
The resume screening process is still very manual with recruiters using filtering to parse through resumes.
02.
Technical skills tend to be the primary focus when recruiters are screening through resumes.
03.
Recruiters' strategy for being unbiased is making sure that all resumes that are submitted are screened.
04.
Job titles and skills can vary in wording within resumes even if they are describing the same thing.
05.
The way in which recruiters screen resumes can vary with some paying attention to different categories over others even if it is for the same position.
Solution
The primary deliverable of this project is a web-based candidate screening platform that consists of the following components:
1. Job Description Analysis: A RAG-LLM that extracts that extracts features like skills, experience, industry preferences, and education from job descriptions.
2. Resume Matching: Resumes also undergo entity extraction and are then matched against job requirements. Candidates are bucketed by match score into: Highly Qualified, Moderately Qualified and Potential Matches .
3. Transparent and Flexible Scoring: Breakdown of scoring criteria weightage for different features, which are completely adjustable. Users can view the specific requirements that a candidate matches or does not match, along with access to their resumes.
4. Interaction Tracking: The platform tracks user interactions, such as the resumes they view or interact with, to identify potential biases. A pop-up displays the gender and ethnic distribution of the screened candidates.
5. User Interface: Simple UI for displaying candidate dashboard and setting job requirements.provided to display the results of the candidate matching process and allow users to control the weights and priorities.
TECHNOLOGY USED:
-
RAG-LLM: It is used to retrieve and incorporate relevant information from corpus of documents to generate more informed and context-aware responses. We used this for entity extraction from documents
-
Machine Learning: ML techniques like NLP and other text processing techniques like n-gram matching, were used to rank and score candidates
-
Backend: FastAPI for building the API and server-side logic.
-
Frontend: Flask for developing the user interface.
-
Database: Vector DB for storing job descriptions, resumes, and SQLite for storing structured profiles for both.
-
Libraries/Frameworks: Relevant libraries like LangChain, Hugging Face, nltk, etc.
DEMO:
RESULTS:
We successfully developed a prototype of the candidate screening platform, demonstrating the feasibility to use AI to make the screening process objectively based on merit, promoting unbiased evaluation. Initial testing showed promising results in reducing biases and providing a transparent, data-driven evaluation approach. However, further refinement is required to address shortcomings, while having continuous monitoring and feedback to ensure fairness during real-world deployment.
In collaboration with:
Douglas Tan Lew: UX Researcher
Jin Lee: Project Manager
Jay Kuo: Data Scientist / Analyst
Acknowledgements:
We express our sincere gratitude to Included for providing us the opportunity to work on this capstone project. We are especially thankful to Chandan Golla from Included, who encouraged us to learn and experiment, guided the project, and provided valuable direction. Special thanks to Prof. Sturman for his invaluable guidance, advice, and support throughout the project. We also extend our appreciation to our TA, Bingbing Wen, for her dedicated support, insightful discussions, and constructive feedback during our status reports.
To learn about all the cool stuff that Included does, check out their website!