Abstract
Organisations often struggle to identify the causes of change in metrics such as product quality and delivery duration. This task becomes increasingly challenging when the cause lies outside of company borders in multi-echelon supply chains that are only partially observable. Although traditional supply chain management has advocated for data sharing to gain better insights, this does not take place in practice due to data privacy concerns. We propose the use of explainable artificial intelligence for decentralised computing of estimated contributions to a metric of interest in a multi-stage production process. This approach mitigates the need to convince supply chain actors to share data, as all computations occur in a decentralised manner. Our method is empirically validated using data collected from a real multi-stage manufacturing process. The results demonstrate the effectiveness of our approach in detecting the source of quality variations compared to a centralised approach using Shapley additive explanations.
The Challenge of Multi-Echelon Supply Chains
Modern supply chains are complex, multi-echelon systems where products traverse multiple independent organizations before reaching the final consumer. When systemic issues arise—such as a sudden drop in product quality or unexpected delivery delays—pinpointing the root cause is notoriously difficult. A defect might originate from a Tier 3 raw material supplier, a Tier 1 sub-assembly plant, or the final manufacturer's own calibration.
While traditional supply chain management suggests centralizing data to gain visibility, this is rarely feasible. Companies treat their process data as highly sensitive intellectual property, fearing that sharing it could lead to competitive disadvantages or price negotiations. Consequently, these supply chains remain "black boxes," hindering efficient problem-solving.
A Decentralized Framework for Contribution Estimation
To address these privacy concerns, this paper introduces a decentralized framework using Explainable AI (XAI). Instead of pooling raw data, the method leverages Deep Ensembles for uncertainty estimation to identify influential suppliers.
The process follows a structured protocol:
- The Call: The lead manufacturer provides the "metric of interest" (e.g., final quality scores linked to Part IDs) to all participating suppliers.
- Local Computation: Each supplier trains a local ensemble of neural networks on their private data to predict the provided metric.
- Uncertainty Scoring: Suppliers calculate their model's total uncertainty (combining epistemic and aleatoric uncertainty).
- The Reply: Suppliers return only a single value: their total uncertainty score.
- Ranking: The manufacturer ranks these scores. A lower uncertainty score indicates that the supplier's data has a higher predictive power—and thus a greater influence—over the final outcome.
To filter out irrelevant data, the framework includes a "pure noise baseline." If a supplier's uncertainty score does not outperform a model trained on random noise, they are excluded as a significant contributor.
Empirical Validation
The researchers validated this approach across four datasets:
- Manufacturing Case Study: Real-world data from a multi-stage production line in Detroit (provided by Liveline Technologies), featuring machine parameters like RPM, temperature, and material properties.
- UCI Benchmarks: Three tabular datasets (Diabetes risk, Red Wine Quality, and White Wine Quality) were used to simulate independent data sources by grouping features.
Key Results and Insights
The core innovation lies in the realization that uncertainty serves as a proxy for influence. A supplier's ability to reduce predictive uncertainty about the final outcome is a mathematically sound indicator of their contribution.
The decentralized method was benchmarked against SHAP (SHapley Additive exPlanations), the industry standard for centralized feature importance.
- Consistency: The decentralized rankings matched the centralized SHAP rankings across all datasets.
- Accuracy: In the manufacturing case study, the method correctly identified "Company F" as the primary driver of quality variance, aligning with the ground truth established by the centralized model.
- Privacy: By sharing only a single high-level uncertainty score, the method avoids the privacy risks associated with sharing raw data or model weights.
Limitations and Future Outlook
A primary limitation is multicollinearity. When two suppliers' processes are highly correlated (e.g., sharing the same environmental conditions), the model may struggle to distinguish the true driver.
Future work suggests using this uncertainty-based ranking as a "pre-screening" tool for more complex collaborations like Federated Learning. By identifying the most relevant contributors first, companies can focus their resources on the most impactful data partnerships.
Summary updated