Employee Attrition Prediction and Retention Policy Simulation Using Explainable ML

Actionable retention recourse for HR analytics: for every employee at risk of leaving, the system searches every allowed combination of changes to MUTABLE factors only - overtime off, a 5-20 % raise, business travel one step lower, stock options +1/+2, two extra training sessions, a promotion - within plausible bounds (raise + promotion at most +35 %), prices each one and reports the cheapest change that brings the cross-fitted predicted probability of leaving below 0.30. A retention-budget optimiser (multiple-choice knapsack over employees x action bundles, solved exactly as a MILP) decides who gets which action within a budget and is compared with the usual 'top-risk first' rule and a random order. Prediction (logistic regression, random forest, LightGBM; repeated nested cross- validation), SHAP explanations and a fairness check by gender, age band and marital status complete the study. The IBM HR Attrition data is a FICTIONAL sample created by IBM: the results demonstrate the method and are not evidence about real workforces; recourse is a model-based association, not a causal effect.

Category: HR analytics

Master of Business Administration Project MBA03 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.