A SOFT VOTING MECHANISM-BASED RANDOM FOREST MULTI-MODEL ENSEMBLE APPROACH FOR EMPLOYMENT STATUS PREDICTION

Abstract

Employment constitutes the cornerstone of people's livelihoods and the foundation of development. Presently, structural employment contradictions remain prominent in China, rendering the achievement of high-quality, full employment the foremost objective of current socio-economic development. Conventional single-model forecasting approaches exhibit limitations when handling high-dimensional, complex employment data, including insufficient generalisation capabilities and restricted capture of feature interactions. To address this, this study constructs a multi-model fusion employment status prediction model based on random forests, thereby enhancing predictive accuracy and stability. The study utilised a dataset comprising 4,980 employment samples, encompassing 57 raw features. Through data cleansing, Pearson correlation analysis, and feature extraction methods based on XGBoost and AUC cross-validation, 12 key feature variables were ultimately selected. Building upon this foundation, a soft voting ensemble strategy was employed to fuse three major prediction models—random forest, SVM, and LSTM—into a novel ensemble learning model. The predictive performance and stability of both the traditional trio of models and the new ensemble model were evaluated. Results indicate: The ensemble model achieved an accuracy of 85.49%, with recall and F1 scores of 97.34% and 93.21% respectively, outperforming each individual model. It effectively synergises the strengths of multiple models, enhancing adaptability to complex employment data and improving prediction robustness. This research provides a scientifically sound and effective ensemble learning approach for employment status prediction, offering practical value in supporting government precision policy-making and optimising employment services.