員工離職問題為企業人力資源管理之重要議題，過高之離職率不僅增加企業之招募與培訓成本，亦可能影響組織績效與營運穩定性。因此，如何有效辨識影響員工離職之關鍵因素並建立準確之預測模型，已成為企業關注之核心課題。本研究以 IBM HR Analytics 員工資料集為研究對象，透過系統化分析流程，整合資料前處理、特徵工程與機器學習方法，建構員工離職精準預測模型。 在研究方法上，首先進行資料前處理，包括排除冗餘變數、變數轉換與標準化，以提升資料品質。接著透過主效應分析、共線性檢定與相關性分析，篩選出具代表性之關鍵變數，並進一步探討變數間之交互作用，以深入理解員工離職之影響機制 。在模型建構方面，本研究採用多種機器學習方法進行比較，包括 KNN、決策樹、邏輯斯迴歸、支持向量機、單純貝式分類器、多層感知器、隨機森林、AdaBoost、Bagging 及 GDBT 等模型，以評估不同方法之預測效能。 此外，針對原始資料中離職樣本比例偏低之問題，本研究導入生成對抗網路（GAN）進行資料擴增，使離職與留任樣本達到平衡，進而提升模型對少數類別之辨識能力。分析結果顯示，經GAN資料平衡後，各模型在 Accuracy、Recall、F1 及 AUC 等指標上均有顯著提升，其中以 GDBT 與 Bagging 模型表現最佳，其預測準確率可達到91.9%。 本研究建立一套具高預測能力之離職預測模型，亦透過特徵工程分析揭示影響員工離職之關鍵因素與其交互作用關係，進一步提供企業在人力資源管理與人才留任策略上之決策依據。;Employee attrition is a critical issue in human resource management， as excessive turnover not only increases recruitment and training costs but also adversely affects organizational performance and operational stability. Therefore， identifying key factors influencing employee attrition and developing accurate predictive models have become essential concerns for organizations. This study employs the IBM HR Analytics dataset as the research foundation and proposes a systematic analytical framework that integrates data preprocessing， feature engineering， and machine learning techniques to construct a precision prediction model for employee attrition. In terms of methodology， data preprocessing is first conducted， including the removal of redundant variables， variable transformation， and standardization， to enhance data quality. Subsequently， main effect analysis， multicollinearity diagnostics， and correlation analysis are applied to identify significant variables. Furthermore， interaction effects among variables are explored to provide a deeper understanding of the mechanisms underlying employee attrition. For model development， multiple machine learning algorithms are implemented and compared， including KNN， Decision Tree， Logistic Regression， Support Vector Machine (SVM)， Naive_Bayes (NB)， Multilayer Perceptron (MLP)， Random Forest (RF)， AdaBoost， Bagging， and Gradient Boosting Decision Tree (GDBT)， to evaluate their predictive performance. In addition， to address the issue of class imbalance in the original dataset， where attrition cases are underrepresented， this study incorporates a Generative Adversarial Network (GAN) for data augmentation. By generating synthetic samples of the minority class， the dataset is balanced， thereby improving the model’s ability to identify attrition cases. The results indicate that， after data balancing of GAN， all models exhibit significant improvements in performance metrics， including Accuracy， Recall， F1-score， and AUC. Among them， the GDBT and Bagging models achieve the best performance， with prediction accuracy reaching up to 91.9%. This study not only develops a highly accurate attrition prediction model but also reveals key influencing factors and their interaction effects through feature engineering. The findings provide valuable insights for organizations in human resource management and talent retention strategies.
