CASE-BASED REASONING FOR EMPLOYEE ATTRITION USING KNN AND WORD2VEC

Abstract

Employee attrition and downsizing remain critical challenges for organizations, requiring decision support systems that balance efficiency, fairness, and transparency. Traditional models based solely on numerical attributes often fail to capture the semantic richness of unstructured employee data, limiting both predictive accuracy and interpretability. To address this gap, this study proposes a hybrid Case-Based Reasoning (CBR) framework that integrates K-Nearest Neighbors (KNN) with Word2Vec embeddings and Manhattan distance. Using a benchmark dataset of 1,470 employees with 34 attributes, the model was evaluated across multiple configurations, including numerical-only KNN and text-enhanced variants with Word2Vec, TF-IDF, and FastText. The results demonstrate that incorporating textual data via Word2Vec significantly enhances classification quality. Specifically, the proposed KNN–Word2Vec–Manhattan model achieved an accuracy of 0.85, representing a 0.01 improvement over the numerical baseline. Precision increased by 0.08 (from 0.79 to 0.87), recall improved by 0.01 (from 0.84 to 0.85), and F1-score rose by 0.02 (from 0.77 to 0.79). While these improvements are modest, they confirm that semantic encoding of textual attributes meaningfully contributes to more robust and interpretable predictions. The findings suggest that hybrid models combining structured and unstructured data offer a promising pathway toward explainable, data-driven HR decision support systems. Future work should investigate richer and domain-specific textual sources to further enhance semantic representation and predictive performance.