Explainable dual LSTM-autoencoders with exogenous features for anomaly detection and supply chain forecasting

Area of Science:

• Supply Chain Management
• Artificial Intelligence
• Machine Learning

Background:

• Retail systems face increasing complexity, necessitating advanced AI for optimal management.
• Traditional statistical models struggle with nonlinear patterns, seasonality, and external factors in retail data.
• Accurate demand prediction and anomaly detection are crucial for supply chain efficiency, cost reduction, and customer satisfaction.

Purpose of the Study:

• To propose a novel dual-head AI system for simultaneous demand forecasting and anomaly detection in retail supply chains.
• To enhance predictive accuracy and robustness against outliers through specialized network architectures and feature engineering.
• To provide model transparency and interpretability for better understanding of forecasting and anomaly detection logic.

Main Methods:

• A dual-head system combining Long Short-Term Memory (LSTM) networks for forecasting and Autoencoders for anomaly detection.
• Implementation of feature engineering and selection techniques to capture temporal dependencies and structural patterns.
• Empirical evaluation on the M5 Forecasting dataset using error-based, accuracy-based, and prediction interval metrics.

Main Results:

• The proposed model achieved a 9.4% relative improvement over baseline deep learning models.
• Demonstrated lower error rates with an RMSE of 10.84 and robust interval coverage with an MPIW of 5.2.
• Model transparency was achieved using Saliency maps, SHAP values, and LIME explanations.

Conclusions:

• The dual-head AI system significantly improves demand forecasting and anomaly detection in retail supply chains.
• The integration of LSTM and Autoencoders, coupled with feature engineering, enhances model performance and reliability.
• Explainable AI techniques provide valuable insights into model decision-making processes for supply chain management.