Abstract:This paper views hiring as a contextual bandit problem: to find the best workers over time, firms must balance exploitation (selecting from groups with proven track records) with exploration (selecting from under-represented groups to learn about quality). Yet modern hiring algorithms, based on supervised learning approaches, are designed solely for exploitation. Instead, we build a resume screening algorithm that values exploration by evaluating candidates according to their statistical upside potential. Using data from professional services recruiting within a Fortune 500 firm, we show that this approach improves the quality (as measured by eventual hiring rates) of candidates selected for an interview, while also increasing demographic diversity, relative to the firm's existing practices. The same is not true for traditional supervised learning based algorithms, which improve hiring rates but select far fewer Black and Hispanic applicants.
Together, our results highlight the importance of incorporating exploration in developing decision-making algorithms that are potentially both more efficient and equitable.
What problem does this paper attempt to address?
The problem this paper attempts to address is: how to improve the quality of candidates in the recruitment process while increasing diversity without sacrificing efficiency. Specifically, the authors focus on the fact that modern recruitment algorithms (based on supervised learning methods) can improve recruitment quality but often result in fewer interview opportunities for minority and female candidates. Therefore, the paper proposes a new algorithm—the Contextual Bandit algorithm, which aims to balance "exploitation" (selecting from groups with a successful track record) and "exploration" (selecting from underrepresented groups to learn about their quality) through exploratory choices, thereby achieving dual improvements in recruitment quality and diversity.
### Background of the Paper
1. **Fairness and Efficiency Issues in Recruitment**:
- Despite many companies implementing diversity, equity, and inclusion programs, there is still bias against minority and female candidates in the recruitment process.
- While supervised learning algorithms can improve recruitment quality, they often reduce interview opportunities for minority and female candidates.
2. **Limitations of Existing Recruitment Algorithms**:
- Modern recruitment algorithms are primarily based on supervised learning methods, which focus only on predicting the best candidates from existing data, ignoring the importance of exploring new groups.
- This approach can lead to fairness issues in the recruitment process, as the algorithms may favor historically well-performing groups and overlook potentially excellent but underrepresented groups.
### Methodology of the Paper
1. **Data Source**:
- The authors used professional service recruitment data from a Fortune 500 company, studying recruitment decisions for high-skilled positions such as consulting, financial analysis, and data science.
2. **Algorithm Comparison**:
- **Supervised Learning Model (SL)**: Predicts the recruitment potential of candidates based on existing training data and selects the candidates most likely to be hired.
- **Contextual Bandit Model (UCB)**: Considers not only the predicted potential of candidates but also exploratory choices, selecting candidates with higher uncertainty to learn more about them.
### Main Findings
1. **Recruitment Quality**:
- Both algorithms can improve the quality of candidates selected for interviews, but the UCB model performs better in this regard.
- Through various evaluation methods (direct observation, inverse probability weighting, instrumental variable strategy), the authors found that candidates recommended by the UCB model had a higher probability of being hired.
2. **Diversity**:
- The Supervised Learning Model (SL) significantly reduced interview opportunities for Black and Hispanic candidates, from 9.4% to 4.2%.
- In contrast, the UCB model significantly increased interview opportunities for underrepresented minorities, from 9.4% to 24.3%.
### Conclusion
- The study shows that by introducing exploratory choices, recruitment algorithms can improve diversity and recruitment quality without sacrificing efficiency.
- This finding has important implications for other scenarios that require evaluating applicant quality (such as promotions, credit scoring, loan approvals, university admissions, investment, and research funding allocation), indicating that exploratory algorithms can overcome the limitations of historical data and make more comprehensive and forward-looking assessments.
### Significance
- The study emphasizes the importance of incorporating exploratory choices in algorithm design, which is significant for promoting fairness and efficiency in the labor market.
- Through this approach, companies can better identify and attract excellent talent from diverse backgrounds, thereby improving overall recruitment outcomes.