An AI-Driven Predictive Model Using Decision Trees for Detecting Gender Bias in HR Practices and Recommending Policy Interventions

Gender bias, HR analytics, Decision tree, Machine learning, Pay equity, Explainable AI, Fairness in AI, People analytics, Policy recommendation, Workplace gender equality

Despite the prevalent legislative frameworks and diversity efforts, gender prejudice in organisational Human Resource (HR) practices remains a quantifiable and impactful type of inequality in the workplace. This article suggests an AI-based predictive system based on Decision Tree classifiers to actively identify gender bias in five HR practice areas: pay equity, promotion, access to training, employee engagement, and attrition. A new two-step modelling model is proposed: Stage 1 (Reverse Modelling): with the help of HR outcome variables as predictors, the gender of the employees can be predicted which statistically demonstrates the presence of systemic bias; Stage 2 (Forward Modelling): the gender and the HR features are used as predictors to predict an artificial bias label, which indicates the circumstances under which bias appears. This framework is trained on HR_Analytics.csv (1,480 records) and tested on three independent datasets of 22,119 records of employees overall. The experimental findings show reverse model AUC of up to 62.4, which proves that there is a significant gender cue in HR results. Forward modelling has 100 percent accuracy and F1 score on all the dataset. Among the key results are a pay gap of -7.1% in Sales departments established in two independent datasets, deficits of 2.04-2.18 sessions per year of training in Finance and Marketing, and delays in promotions of 0.51-0.57 years in case of female employees in HR and Technology functions. The output of decision trees is converted into specific HR intervention, which is automatically generated into a rule-based policy recommendation engine. Comparative analysis with six other algorithms shows that the Decision Tree is the only algorithm to provide state of the art predictive accuracy and fully interpretable results as well as auto-generated policies, which are not available with black-box ensemble techniques.