Interpretable CNN-Attention Hybrid Framework for Spatiotemporal Feature Engineering in Youth Employment Market Trend Prediction

Abstract

This research propose an interpretable hybrid neural-temporal framework for youth employment trend prediction that integrates dilated convolutional neural networks (CNNs) with self-attention mechanisms to extract and analyze spatiotemporal features from multivariate employment indicators. The framework addresses the dual challenges of capturing multi-scale temporal dependencies and providing policy-actionable insights, which are critical for understanding complex labor market dynamics. The methodology combines a dilated CNN architecture to isolate local patterns such as seasonal fluctuations and abrupt shocks, followed by a modified self-attention mechanism that dynamically weights features and time steps to enhance interpretability. Furthermore, a gating mechanism derives time-aggregated feature importance scores, enabling recursive refinement of high-impact variables during preprocessing. The proposed method interfaces with conventional modules through robust median-based normalization and attention-guided feature selection, which employs LASSO regularization to prioritize influential predictors. Implemented with TensorFlow/Keras and optimized for GPU acceleration, the framework handles high-resolution data efficiently while maintaining transparency in decision-making. Experiments demonstrate its superiority over traditional ARIMA or RNN-based approaches, particularly in scenarios requiring both accuracy and interpretability. The results highlight its potential as a tool for policymakers to identify critical drivers of youth employment trends, thereby supporting targeted interventions and long-term labor market planning.