Study of the influence of a biased database on the prediction of standard algorithms for selecting the best candidate for an interview

TL;DR

The study addresses how biased training data can propagate discrimination in AI-driven recruitment by simulating external discrimination and internal self-censorship across multiple candidate-ranking scenarios. It evaluates five classical algorithms—logistic regression, logistic regression with AIC, L-nearest neighbors, multilayer perceptron, and SVM—on both full (X,Y,Z) and anonymised (X,Z) data, comparing learning from perfect versus biased rankings. Key findings show that predictive quality degrades when training data are biased, while anonymisation can partially mitigate bias when discriminatory features are weakly correlated with job-relevant features; more complex models do not consistently outperform simpler logistic approaches, and LNN may offer gains in certain low-bias conditions. The work provides a practical framework for assessing bias in recruitment AI and informs design choices around data anonymisation and algorithm selection, with implications for cleaner, fairer hiring pipelines.

Abstract

Artificial intelligence is used at various stages of the recruitment process to automatically select the best candidate for a position, with companies guaranteeing unbiased recruitment. However, the algorithms used are either trained by humans or are based on learning from past experiences that were biased. In this article, we propose to generate data mimicking external (discrimination) and internal biases (self-censorship) in order to train five classic algorithms and to study the extent to which they do or do not find the best candidates according to objective criteria. In addition, we study the influence of the anonymisation of files on the quality of predictions.

Figures (25)

• Figure 1: Schematic representation of the notations for $n$ recruitment methods. Each recruitment method $i$ is composed of $N_d$ profiles (one by row in the central box) and each profile is associated at an answer to know if it has the job or not (center right matrix).
• Figure 2: Schematic representation of the biased ranking. Starting from the threshold $S$, profiles are separated into two categories: discriminated (bottom) and non-discriminated (top). The good profiles and variables $\overline{X}_{i,j}$ are used to start the ranking, while the bad profiles and variables $\overline{Y}_{i,j}$ are used to end the ranking.
• Figure 3: Schematic representation of a neuron (left) and a neural network with a hidden layer containing $p$ neurons (right).
• Figure 4: Schematic representation of a scatter plot with two labels ($\textcolor{blue}{\blacksquare}$ and $\textcolor{red}{\bullet}$) where the boundary is along a derivable curve (left) and its projection by a $\Phi$ function so that the separation is according to a hyperplane (right). The boundary is symbolised by a dashed line.
• Figure 5: Representation of logistic regression results for the self-censorship scenario, where each point has as its $x$-axis the average rate of correct classifications if the algorithm were to train on perfect classification, and as its $y$-axis the biased classification according to discrimination case (columns), $\alpha$ correlation (rows) and file type (in turquoise for complete files and saumon for anonymized files). The black line represents $y=x$ and ellipses at 95% have been added.