Graph Spatio-Temporal Networks for Manufacturing Sales Forecast and Prevention Policies in Pandemic Era

Area of Science:

• Data Science
• Industrial Engineering
• Econometrics

Background:

• COVID-19 pandemic significantly impacted global manufacturing supply chains due to sourcing, globalization, and inventory dynamics.
• Spatial-temporal models and Graph Convolutional Networks (GCNs) are increasingly used for analyzing complex time-series data.
• Existing Attention-Adjusted Graph Spatio-Temporal Networks (AGSTN) face challenges with small, non-sensor datasets, particularly convergence issues.

Purpose of the Study:

• To propose and evaluate novel variants of AGSTN tailored for non-sensor data.
• To enhance AGSTN performance by incorporating data augmentation and regularization techniques.
• To address the convergence problem in AGSTN when applied to limited datasets.

Main Methods:

• Development of several AGSTN variants.
• Application of data augmentation and regularization techniques: edge selection, time series decomposition, and prevention policies.
• Empirical validation on worldwide manufacturing industry data from the pandemic era.

Main Results:

• The proposed AGSTN variants demonstrated significant improvements in prediction performance.
• Mean Squared Error (MSE) was reduced by at least 20% compared to baseline models.
• The enhanced methods effectively addressed the convergence problem observed in AGSTN.

Conclusions:

• The modified AGSTN variants offer a robust solution for analyzing spatio-temporal correlations in non-sensor manufacturing data.
• Data augmentation and regularization are crucial for improving AGSTN performance and stability.
• This research provides valuable insights for understanding and predicting manufacturing industry dynamics during crises.