Authors: Alka G. Saraf, J Rathnamala
Abstract: Attrition and employee engagement remain among the most pressing concerns related to human capital, directly impacting organizational effectiveness, but current techniques for predicting such outcomes rely solely on limited survey data. In this paper, we propose an end-to-end multimodal workforce analytics solution that utilizes structured HR information (employee demographics, performance evaluation metrics, remuneration), semi-structured textual information (exit interviews, management feedback), and behavioral time-series data (usage statistics of internal communication platforms and badge access logs). The proposed predictive model uses multimodal transformer with cross-modal attention techniques to jointly forecast the likelihood of employee attrition (binary classification task, AUROC = 0.89) and their overall engagement (regression task, MAE = 0.31). Tested on data collected over 18 months for 8,472 employees at a multinational IT company, our method discovers distinctive behavioral indicators, with the decrease in collaboration entropy and higher activity outside regular hours predicting attrition 12 weeks in advance. By combining NLP techniques for parsing exit interviews, we discovered that "career development opportunities" and "management competency" were the top textual predictors of leaving the job.
DOI: https://doi.org/10.5281/zenodo.20233561
