Deep learning framework for interpretable supply chain forecasting using SOM ANN and SHAP.

Area of Science:

• Artificial Intelligence
• Supply Chain Management
• Data Science

Background:

• Industry 4.0 presents complex, data-rich supply chains with real-time uncertainties.
• Traditional supply chain management (SCM) struggles with these dynamic environments.
• AI-guided predictive modeling is essential for modern SCM.

Purpose of the Study:

• To propose a deep learning (DL) framework for enhanced supply chain shipping time prediction and delivery risk assessment.
• To improve the accuracy and robustness of SCM predictions compared to conventional methods.
• To ensure model interpretability for end-user comprehension and trust.

Main Methods:

• A deep learning framework combining Self-Organizing Maps (SOMs), Principal Component Analysis (PCA), and Artificial Neural Networks (ANNs).
• Application of the SOM+ANN model to the DataCo Smart Supply Chain dataset for predicting shipping duration and delivery risk.
• Utilized SHAP (Shapley Additive exPlanations) for model interpretability.

Main Results:

• The SOM+ANN model significantly outperformed conventional Machine Learning (ML) models like Random Forest, XGBoost, and Decision Tree.
• Achieved R² of 0.92, RMSE of 0.936, and MAE of 0.8459 for shipping duration prediction.
• Demonstrated 96% accuracy and 96.22% F1-score for delivery risk classification on dataset 1, and 89.65% accuracy on dataset 2.

Conclusions:

• The proposed SOM+ANN framework offers a more accurate and robust solution for SCM prediction tasks.
• Transparent AI methodologies, like SHAP, enhance the understanding and adoption of AI in SCM.
• The framework improves SCM operations, resilience, and decision-making capabilities.