Generating Logically Consistent Synthetic Supply Chain Data with LLM-Driven Knowledge Graph Reasoning

Technical Summary: Generating Logically Consistent Synthetic Supply Chain Data with LLM-Driven Knowledge Graph Reasoning

Problem Statement
Synthetic data is increasingly used to address data scarcity and privacy barriers in supply chain analytics. However, existing tabular generative models (e.g., GANs, diffusion models, LLM-based generators) are primarily optimized for statistical fidelity—reproducing marginal distributions and correlations—rather than preserving the operational logic that governs real-world supply chain processes. Consequently, generated records often violate fundamental constraints, such as temporal orderings (e.g., delivery dates preceding order dates), mathematical dependencies (e.g., totals not equaling price $\times$ quantity), and hierarchical taxonomies (e.g., cities mapping to incorrect countries). The authors argue that for synthetic data to support operational simulation and decision-making, it must adhere to the "physics" of supply chain data: the structured, domain-specific rules that make a record operationally plausible. Current methods fail to explicitly discover or enforce these schema-specific constraints, leading to data that is statistically realistic but logically invalid.

Methodology: The TabKG Framework
The paper introduces TabKG, a knowledge-graph-guided framework designed to generate logically consistent synthetic tabular data. The approach operates through five distinct stages:

1. Metadata Serialization: Column names and descriptions from the input table schema are serialized into natural language prompts.
2. Multi-LLM Ensemble Graph Construction: A Column Relationship Knowledge Graph (CR-KG) is constructed to represent inter-column dependencies. Instead of relying on a single model, TabKG employs an ensemble of $K$ Large Language Models (LLMs) with majority voting. Each LLM proposes candidate edges (relationships) between columns, categorized by type: Hierarchical (H), Mathematical (M), Temporal (T), or Semantic (S). An edge is retained only if proposed by a majority of the models.
3. Data-Driven Validation and Pruning: The candidate graph is validated against the real dataset to eliminate hallucinated or unsupported edges. Type-specific validators compute a confidence score $\sigma(e)$ for each edge:
   • Hierarchical: Checks functional dependency (does the source uniquely determine the target?).
   • Mathematical: Evaluates formula accuracy on real rows.
   • Temporal: Checks for ordering violations.
   • Semantic: Evaluates conditional rule satisfaction rates.
     Edges with scores below a threshold (default $\theta = 0.90$) are pruned, resulting in a validated CR-KG.
4. Graph-Guided Compression and Generation: The validated graph is converted into a Directed Acyclic Graph (DAG) to identify a minimal set of independent columns ($V_{keep}$). Dependent columns are compressed out. A latent score-based diffusion model is trained only on the independent columns to generate synthetic compressed data ($X'_c$).
5. Knowledge Graph-Guided Decompression: The full synthetic dataset is reconstructed by deterministically applying the validated relationships to the generated independent columns. Dependent columns are restored using lookup tables (for hierarchical/semantic), closed-form formulas (for mathematical), or additive time offsets (for temporal). This ensures logical consistency is enforced by construction rather than learned implicitly.

Key Contributions

• Automatic Discovery of Operational Logic: TabKG is the first method to automatically construct a knowledge-graph representation of inter-column supply chain logic from tabular schemas without manual rule engineering. It combines LLM semantic knowledge with empirical data validation.
• Compression-and-Reconstruction Pipeline: The authors propose a generation pipeline that separates independent and dependent variables. By generating only independent columns via diffusion and deterministically reconstructing dependent ones, the framework guarantees that all validated operational rules are satisfied in the output.
• Comprehensive Evaluation: The framework is evaluated across two industrial datasets (Retail and Purchasing) and two downstream tasks (late-delivery risk prediction and procurement-status classification), assessing fidelity, utility, privacy, and logical consistency.

Experimental Results

• Logic Reasoning: In the task of inferring inter-column relationships from metadata, TabKG (using a same-model ensemble of DeepSeek) achieved an F1 score of 0.897 on the Retail dataset and 0.945 on the Purchasing dataset. This substantially outperformed prompt-only baselines, which ranged from 0.27 to 0.55.
• Logical Consistency: TabKG achieved Hierarchical Consistency Scores (HCS) and Multivariate Dependency Indices (MDI) of 97.84%–98.73%, significantly outperforming state-of-the-art baselines (CTGAN, TabDDPM, TabSyn, GReaT), which often collapsed to near-zero HCS on hierarchical tasks.
• Data Utility: In Train-on-Synthetic, Test-on-Real (TSTR) classification tasks, models trained on TabKG data achieved AUC scores of 71.90 (Retail) and 63.45 (Purchasing), outperforming all other generative baselines and preserving roughly 75–82% of the performance of models trained on real data.
• Fidelity and Privacy: TabKG maintained statistical fidelity comparable to top baselines (e.g., TabSyn) in terms of density estimation and correlation, while achieving strong privacy metrics (Distance to Closest Record and Classifier Two-Sample Test), ensuring no direct memorization of real records.

Significance and Claims
The paper claims that TabKG provides a foundation for moving synthetic data beyond simple machine learning augmentation toward high-fidelity supply chain simulation and operational decision support. By explicitly preserving the "physics" of supply chain operations (temporal orderings, mathematical formulas, and hierarchical rules), TabKG generates data that is not only statistically realistic but also operationally valid. This addresses a critical gap where existing methods produce data that fails in realistic simulation scenarios due to logical inconsistencies. The authors position TabKG as a training-free, schema-agnostic solution that leverages pre-trained LLM knowledge to adapt to diverse supply chain schemas (retail, procurement, manufacturing) without requiring bespoke rule libraries, thereby lowering the barrier for organizations to generate trustworthy synthetic data for simulation and autonomous orchestration in Supply Chain 4.0.