Dynamic Context-Weighted Embeddings: A Novel Approach to Predictive Modelling

Abstract

Employee satisfaction prediction models often struggle to capture the complex, non-linear relationships between compensation and job satisfaction, particularly in heterogeneous organisational contexts. This paper introduces a novel deep learning framework incorporating multiple technical innovations to address these challenges. The proposed approach employs a dual-pathway neural architecture with compensation-specific processing modules to explicitly model the non-linear interactions between compensation factors and other job attributes across diverse organisational settings. A differentiated embedding strategy transforms raw features into rich, context-aware representations, enabling the capture of subtle patterns in employee satisfaction dynamics. The framework integrates a context-sensitive attention mechanism that automatically identifies and weighs relevant features based on organisational characteristics and temporal patterns, alongside a specialised loss function that adaptively emphasises difficult-to-predict cases, improving performance on complex satisfaction patterns. This model demonstrates robust performance across diverse industry settings, handling missing data and class imbalance effectively. Extensive comparative experiments against state-of-the-art methods (LSTM-Attention, GNN-based approaches and traditional ML models) across multiple datasets show significant improvements, with prediction accuracy increasing by 5.2%–8.5%, mean squared error decreasing by 10.3%–15.2% and AUC-ROC metrics improving by 7.8%. Further analysis reveals superior performance in handling temporal dependencies and organisational context variations, with particular strength in predicting satisfaction levels during significant organisational changes.