CLASNet: A Cognitive Load–Aware CNN-LSTM-Attention Framework for Supply Chain Demand Forecasting and Adaptive Human–Computer Interaction

This paper proposes CLASNet, a Cognitive Load–Aware CNN-LSTM-Attention framework for supply chain demand forecasting and adaptive human–computer interaction. Unlike conventional deep learning forecasting approaches that primarily emphasize predictive accuracy while neglecting human cognitive constraints, CLASNet integrates demand prediction and cognitive load modeling within a unified end-to-end architecture to improve both forecasting performance and decision-support efficiency. The proposed framework combines a CNN-LSTM demand feature encoder for capturing local temporal patterns and long-term dependencies, a Cognitive Load Encoder (CLE) for modeling user interaction behaviors, a Cognitive-Aware Attention Fusion (CAAF) mechanism that dynamically adjusts temporal attention weights according to users’ cognitive states, and an Adaptive Interaction Decision Layer (AIDL) that optimizes visualization complexity and information density. Experiments were conducted on a real-world retail supply chain dataset containing three years of demand records and dashboard interaction logs collected from intelligent decision-support systems. Comparative results against ARIMA, LSTM, CNN-LSTM, and CNN-LSTM-Attention baselines demonstrate that CLASNet achieves superior forecasting accuracy, obtaining an RMSE of 14.55, MAE of 9.22, and MAPE of 10.9%, corresponding to an approximately 9.6% RMSE reduction compared with the strongest baseline. Ablation studies further verify the effectiveness of the cognitive load encoder, cognitive-aware attention mechanism, and adaptive interaction module. In addition, user-centered evaluations indicate that the proposed cognitive-aware adaptive interface significantly reduces task completion time and perceived cognitive load during demand analysis tasks. Overall, the results demonstrate that integrating cognitive awareness into deep learning–based supply chain forecasting systems can substantially enhance both predictive performance and human–AI collaborative decision-making efficiency.