Reducing Bias in AI-based Recruitment Systems utilizing ML and NLP Tools by Identifying Gender Gaps in LinkedIn Profiles
M.Sc. Candidate: Vivian Simon
Instructors: Dr. Neta Rabin and Dr. Hila Chalutz Ben-Gal
In recent years, significant efforts have been made to eliminate discrimination in online employee recruitment and selection processes. Systems that incorporate algorithmic decision-making and other forms of profiling may lead to discrimination based on gender, race, religion, nationality, or other socio-demographic features. In modern workforce, the growing practice of social networks embedded job search, such as LinkedIn and Facebook, become major recruitment tools. This provides recruiters with access to personal information which may lead to conscious or unconscious bias.
In this work, we analyze gender differences based on the way male and female candidates describe themselves on LinkedIn. We aim to reveal different socio-demographic factors which correlated with self-presentation differences. For this analysis we utilized a training dataset of 14K candidates' LinkedIn profiles from 34 technological positions shared by Talenya, an AI based recruitment engine, along with addition of artificial skills to each candidate. We apply standard and advanced statistical tests on various socio-demographic and organizational sub-groups extracted from the dataset.
We analyze the distributions of male and female candidates within the LinkedIn profile, which contain data of both textual and numerical features. To study and compare multi-dimensional data distributions, we use a machine learning method named the kernel two-sample test. NLP representations together with the cosine similarity distance are applied to explore and analyze textual self-presentation patterns of male and female candidates.
Our results shed light on various candidate populations which have larger gender differences as they are represented in their LinkedIn profile. Those observations may be incorporated in AI recruitment tools intended to overcome bias applying word modification techniques. Identifying biased sub-groups and specific weaknesses in their LinkedIn profiles, may be beneficial for both candidates, as well as to recruiters, and may assist in promoting their online job ranking and visibility.
This work demonstrates three main contributions. First, theoretical contribution to the domain of AI bias and Human Resources Analytics (HRA). We shed light on new aspects of gender gaps specifically in technological settings based on big-data extracted from LinkedIn profiles. Second, this work presents a methodological contribution to the HRA field and utilization of new tools for the purpose of detecting AI bias. Finally, this work presents a practical contribution by providing a robust decision support tool for recruiters and AI recruitment developers in order to reveal new insights related to detectable gender gaps and respective corrective actions.
Bio: Vivian Simon is a M.sc student at the Department of Industrial Engineering, Tel Aviv University. Vivian has been working on her Master’s thesis under the supervision of Dr. Neta Rabin and Dr. Hila Chalutz Ben-Gal. Vivian's work, focuses on identifying AI gender bias based on male and female textual self-presentation on LinkedIn. Vivian holds a B.Sc. degree in Industrial Engineering and Management from Tel Aviv University.
