PREDICTING EMPLOYEE ATTRITION USING TEMPORAL FUSION TRANSFORMERS: A HYPERPARAMETER-OPTIMIZED DEEP LEARNING APPROACH

Employee attrition is of critical concern to organizations, impacting productivity, workforce planning, and bottom lines. Existing machine learning models falter in capturing the complex temporal dynamics in employee behavior, resulting in poor predictive performance. To overcome this, we suggest a Temporal Fusion Transformer based deep learning approach for employee attrition prediction, which is optimized using Bayesian Optimization. IBM HR Analytics Employee Attrition dataset is utilized, employing extensive data preprocessing such as dealing with missing values, encoding categorical features, and scaling numerical features. The TFT model utilizes multi-head attention, Gated Residual Networks, and variable selection mechanisms to acquire short-term and long-term dependencies. Bayesian Optimization optimizes hyperparameters effectively with reduced computational expense and improved performance. The model offers a 97.5% predictive accuracy, 96.8% precision, 97.2% recall, and an AUC-ROC of 98.1%, significantly better than state-of-the-art machine learning methods available today. Organizations can use it to accurately forecast attrition patterns and adopt retention policies ahead of time, thus minimizing turnover and improving job satisfaction. The research contributes to the field of HR analytics with the development of a new deep learning model dedicated to attrition forecasting. As future research, work will continue with the application of explanation methods, such as SHAP, LIME, and attention visualization, in an effort to support explainability, fairness, and interpretability in AI-based HR decision-making and workforce management.