Deep generative neural network dynamics modeling and employee satisfaction prediction for human resource management

Traditional human resource management has limited ability to capture the temporal evolution of employee satisfaction and to learn informative representations from high-dimensional, multi-source data, thereby constraining the accuracy and interpretability of dynamic prediction. To address this issue, I developed a joint VAE-Neural ODE framework that integrates a variational autoencoder (VAE) with a neural ordinary differential equation (neural ODE). Specifically, the VAE encoder employs the reparameterization trick to learn the latent probability distribution of heterogeneous employee records and generate compact yet informative latent features. On this basis, the neural ODE models the continuous-time evolution of latent states and propagates them through a differentiable ODE solver for temporal forecasting. The decoder then maps the evolved latent state to a continuous employee satisfaction score normalized to the interval [0, 1]. Accordingly, the prediction task was formulated primarily as a regression problem, and model performance was evaluated using mean absolute error (MAE) and root mean squared error (RMSE). For interpretive comparison, the continuous outputs were further discretized using a fixed threshold to derive a binary accuracy metric. Experimental results showed that the proposed method achieves an accuracy of 94.7%, an MAE of 0.031, and an RMSE of 0.052 on the test set, while maintaining 87.2% accuracy in cross-quarter forecasting. These findings indicate that the proposed framework can effectively capture nonlinear satisfaction dynamics and provide quantitative support for data-driven human resource analysis. Moreover, by formulating employee satisfaction prediction as a latent continuous-time dynamical process and evolving hidden states through an ODE flow rather than purely discrete recursive updates, the framework produces smoother latent trajectories and reduces error accumulation over longer forecasting horizons and under irregular observation intervals, which are common in non-uniformly sampled HR records.