Enhanced Time Series Forecasting of Employment Ratio Using Stacked Deep Learning Models

The proposed Stacked BiLSTM model demonstrates enhanced predictive accuracy by effectively capturing both past and future temporal relationships within the sequence data. This comprehensive model provides precise and explainable results for forecasting employment trends, facilitating more informed decision-making in the development of economic and workforce policies. The performance of the model is evaluated using metrics such as an R² of 0.9801, an Mean Squared Error(MSE) of 0.0008, an Mean Absolute Error(MAE) of 0.0225, an Explained Variance Score of 0.9845, and a Pearson Correlation of 0.9924, which explicitly surpasses the performance of GRU and LSTM models. The reliability and clarity of the model are further analysed through visualisation techniques, comprising residual distribution plots and training-validation MAE curves, while SHAP and LIME analyses provide meaningful insights into feature contributions.本研究所提出的堆叠双向长短期记忆网络（Stacked BiLSTM）通过有效捕捉序列数据中的过往与未来时间依赖关系，显著提升了预测精度。该综合模型可针对就业趋势预测生成精准且可解释的结果，助力经济与劳动力政策制定过程中的科学决策。本模型通过多项评估指标验证性能，包括决定系数（R²）达0.9801、均方误差（MSE）为0.0008、平均绝对误差（MAE）为0.0225、解释方差分为0.9845，以及皮尔逊相关系数为0.9924，其性能显著优于门控循环单元（GRU）与长短期记忆网络（LSTM）模型。本研究进一步通过可视化技术分析模型的可靠性与清晰性，具体包含残差分布图与训练-验证MAE曲线；同时借助SHAP与LIME分析，可对特征贡献度获得具有实际意义的解读。