PaperlayerView exact revision on arXiv

Source-linked AI summary

Recovering Business Rules as Cascade Lookups for Enterprise Field Prediction

Changxu Duan

arXiv:2610.34258v1cs.LGcs.AIcs.DBcs.IR

TL;DR

Incomplete or stale enterprise configuration rules leave some order fields to manual selection, while statistical methods may miss specific business-key rules. This paper uses KG-guided key discovery and SQL profiling to build deterministic cascades; they outperform evaluated GNN baselines on all eight SALT-KG fields, with results varying by domain and prediction pattern.

  • Problem

    Manually maintained Access Sequences can become incomplete or outdated, while learned methods may fail to capture field values governed by specific business-key rules.

    Source: §1 · pp. 2–3

  • Method

    An LLM uses KG descriptions and SQL profiling to discover and validate business keys, then codifies selected rules as ranked deterministic cascades.

    Source: §1 · p. 3, §3.3 · p. 8

  • Results

    KG-grounded cascades outperform evaluated GNN baselines on all 8 SALT-KG fields, match GNNs on stable recoverable rules, and underperform when prediction requires temporal graph reasoning.

    Source: §4.2 · p. 9, §6 · p. 16

  • Takeaways & Limitations

    The evaluation suggests domain knowledge assets are valuable for prediction when data follows recoverable configuration rules, while drift analysis can provide actionable diagnostics.

    Source: §5 · p. 14, §6 · p. 16

    The system has been validated offline on historical data and has not yet been deployed in a production ERP environment.

    Source: §5 · p. 16

Abstract

from arXiv · show

Enterprise sales-order systems rely on manually maintained configuration rules to determine fields such as payment terms, shipping points, and sales ownership. When these rules are incomplete or stale, operators must enter values by hand, and downstream errors spread into billing, logistics, and reporting. This paper studies whether missing or outdated rules can be recovered from historical transactions instead of being relearned as opaque statistical patterns. We introduce a knowledge-graph-guided workflow in which an LLM agent proposes candidate business keys, meaning column combinations such as customer and document type, validates them through SQL profiling, and compiles the resulting evidence into deterministic cascade lookups. The LLM is used only during offline rule discovery. At inference time, prediction is a transparent key lookup with no model-serving requirement. Experiments on enterprise sales-order prediction and public relational benchmarks show that cascades are most effective when fields follow stable business rules, while learned graph models remain preferable when prediction depends on temporal or neighborhood patterns. Drift analysis further turns residual errors into diagnostics for configuration maintenance. Code and data: https://github.com/Fireblossom/CascadER.

1 Introduction

Enterprise order fields are determined by business rules, but manually maintained configurations can become incomplete or stale. The paper recovers rules from historical transactions through KG-guided key discovery and SQL validation, then deploys them as deterministic cascades.

1 Introduction

  • Stale or incomplete Access Sequences can leave fields blank or apply coarse defaults, forcing representatives to select values manually.Changing assignments can also make once-valid rules outdated.Source: §1 · p. 2
  • Configured determinism means values follow explicit lookup rules keyed by specific combinations of customer, document type, material, and organizational unit.Existing learning approaches struggle when these specific rules, rather than broad feature correlations or neighborhood similarity, determine high-cardinality fields.Source: §1 · p. 3
  • The workflow uses KG descriptions to propose business keys, validates candidates through SQL profiling, and compiles selected rules into deterministic JSON cascades.The LLM is used offline, avoiding per-order model calls at inference time.Source: §1 · p. 3
  • The paper reports its best results among evaluated baselines on all 8 SALT-KG fields, studies cascade boundaries on two public datasets, and analyzes drift-related residual errors.The eight target fields span order headers and line items.Source: §1 · p. 3, Table 1 · p. 4

2 Related Work

Prior work addresses relational prediction, semantic knowledge, LLM-assisted data tasks, tabular models, and enterprise field prediction. This paper instead uses domain knowledge to guide deterministic rule discovery and lookup generation.

Related work

  • RelBench and RelBench v2 establish relational prediction benchmarks with graph-based baselines, while this method generates deterministic lookup tables instead of learning node embeddings.RelBench v2 adds autocomplete tasks, temporal splits, and stronger GNN and LightGBM baselines.Source: §2 · p. 4
  • Prior KG methods use semantics as embeddings, features, or constraints; this approach uses the KG to guide SQL candidate generation and deterministic rule validation.The workflow operationalizes ontology descriptions and determination dependencies as a search guide.Source: §2 · p. 4
  • Unlike general-purpose LLM data-analysis work, this method targets systematic discovery and codification of deterministic business rules into prediction cascades.It builds on LLM workflows that interleave reasoning and actions, including ReAct.Source: §2 · p. 5
  • TabPFN and ConTextTab face scale constraints for SALT-KG’s 2.3M rows and targets with up to 589 classes; JSON lookups require no model serving at inference.ConTextTab requires 80 GB GPU memory with a context window of 8,192 rows and bagging for larger tables.Source: §2 · p. 5
  • Prior enterprise benchmarks and KG embeddings inform this work, which applies KG-guided SQL profiling and uses drift analysis to identify changes transaction history may miss.SALT-KG reported modest MRR improvements from KG-derived embeddings.Source: §2 · p. 5

3 Method

The method defines temporal relational autocomplete, discovers business keys through iterative KG-guided SQL profiling, and codifies selected mappings as ranked lookup cascades. At runtime, it uses lightweight lookup files rather than model inference.

3.1 Problem Formulation

  • The task predicts each transaction’s target field from order-entry inputs, using records before a temporal cutoff for training and records at or after it for evaluation.A composite key selects input columns; lookup values are ranked by empirical frequency among training records, subject to minimum support.Source: §3.1 · p. 5
  • A cascade tries lookup tables from specific keys to general fallbacks, ending at the global target distribution when no earlier key matches.The objective is a short cascade with accurate, high-coverage, temporally valid, auditable predictions.Source: §3.1 · p. 6

3.2 Rule Discovery (Explore)

  • The LLM agent iteratively proposes keys from KG hints or error analysis, profiles them with SQL, and refines the cascade using validation accuracy, coverage, and minimum support.Candidates include single-entity keys, KG-suggested pairs, and higher-order composites; escalation requires an accuracy gain above a threshold.Source: §3.2 · pp. 6–7
  • For CustomerPaymentTerms, SoldToParty plus SalesDocType reached 82.9% accuracy with 95.3% coverage and was promoted over the 80.0%-accuracy customer-only fallback.The global target distribution is appended for uncovered rows.Source: §3.2 · p. 8

3.3 Cascade Codification (Build)

  • Each cascade level stores target values ranked by training support, and inference returns the ranked list from the first matching level.The CustomerPaymentTerms cascade uses SoldToParty plus SalesDocType, then SoldToParty, then the global distribution.Source: §3.3 · p. 8

3.4 Inference (Predict)

  • The compiled JSON lookup files need no LLM, KG, or model inference at runtime; cascades range from 88 keys for Plant to 174K for HeaderIncoterms.The KG narrows key search, while deterministic validation—not the LLM—selects accepted rules.Source: §3.4 · pp. 8–9

4 Experiments

Experiments compare cascade lookups with learning baselines across enterprise and public relational tasks, then examine how key selection, temporal drift, and case-level errors shape their performance. Cascades lead on all eight SALT-KG fields, while public-task results show limits when prediction depends on temporal or neighborhood patterns.

4.1 Setup

  • Evaluation uses 1.9M training and 403K test rows across 8 SALT-KG tasks, plus two RelBench v2 tasks, with temporal cutoffs preventing label leakage.Accuracy uses the top-ranked value; MRR uses the true value’s reciprocal rank, with MODE serving as the lookup-entry estimator.Source: §4.1 · p. 9

4.2 SALT-KG Results

  • The final cascade outperforms the evaluated GNN on all 8 SALT-KG fields; CustomerPaymentTerms reaches 82.9% accuracy versus 37.5%, and SalesGroup reaches 70.0% versus 15.8%.For near-deterministic Plant, ShippingPoint, and SalesOffice mappings, both methods exceed 98%.Source: §4.2 · pp. 9–10, Table 3 · p. 10

4.3 Semantic Key-Selection Ablation

  • Semantic key selection helps most on cross-entity fields, yields 2%–4% gains on composite-key fields, and adds little on already-simple fields.The ablation compares schema-only MODE with a restricted KG-guided cascade before the main method’s error-driven refinement and fallback optimization; it is not exhaustive across rule-mining methods.Source: §4.3 · pp. 10–11, Table 4 · p. 11

4.4 Generalization Study

  • On rel-arxiv, recency-weighted author-category MODE reaches 50.3% test accuracy, while on rel-stack, per-user badge MODE predicts bronze for every activity level.The tasks use temporal and entity-level keys rather than the full KG-guided discovery workflow; rel-arxiv validation accuracy is 52.8% versus GNN’s 52.6%.Source: §4.4 · pp. 11–12

4.5 Drift Root-Cause Analysis

  • Header and item Incoterms co-drift 98%–99%, while Plant→ShippingPoint co-drift is 95% versus 58% in reverse; SalesOffice drifts for only 0.3% of customers.The analysis measures changes in per-customer dominant values across the temporal split for 5,576 customers present in both periods.Source: §4.5 · p. 13, Figure 3 · p. 13
  • Among four non-trivial fields, drift rate and test accuracy have a directionally suggestive linear anti-correlation (slope −0.54 pp per pp drift, R2 = 0.98), based on only four fields.CustomerPaymentTerms has 17.8% drift and 82.9% accuracy; ShippingCondition has 43.8% drift and 69.6% accuracy, and both cascades and GNNs face this ceiling.Source: §4.5 · pp. 13–14

4.6 Error Analysis and Case Studies

  • Historical SoldToParty × SalesDocumentType keys map a standard-order customer to payment term 03 (net 30 days) and correctly predict all test-period orders.Conversely, organizational restructuring can make historical lookup rules obsolete in ways transaction data alone cannot predict.Source: §4.6 · p. 14
  • Plant changes coincide with ShippingPoint changes 95% of the time, but the reverse occurs 58%; configuration owners should review shipping points after plant reassignment.HeaderIncoterms and ItemIncoterms co-drift 98%–99%, a relationship individual field-level audits might miss.Source: §4.6 · p. 14

5 Discussion

The discussion recommends choosing cascades or GNNs according to how well available domain knowledge aligns with observable data. It also frames deployment as a field-specific, auditable process while noting that offline validation and unobserved organizational changes bound the evidence.

When to use cascades in practice

  • Cascades excel with expert knowledge and stable rules, match GNNs without that structure, and underperform when prediction requires temporal reasoning.The discussion recommends GNNs when signals lie in graph neighborhoods or temporal patterns, and drift analysis as an operational diagnostic regardless of method.Source: §5 · p. 14

Practical advantages

  • Cascades provide interpretable composite-key provenance, CPU JSON lookups without model-serving cost, and rapid rebuilds compared with GPU retraining.The reported storage is 12 MB for eight fields versus about 200 MB of model weights; rebuilding takes minutes rather than hours.Source: §5 · p. 15

Prototype-to-production path

  • The evaluation is offline; shadow-mode deployment and operational measurements are required before claiming realized workload reduction.Proposed measurements include acceptance, overrides, corrections, lookup latency, and configuration-owner assessment of drift-alert precision.Source: §5 · p. 15

Ethics and governance

  • Cascade mappings should be audited for protected-attribute correlations, periodically rebuilt, and surfaced as editable suggestions with operator override authority.Predictions log the matched cascade level and composite key for post-hoc analysis and regulatory compliance.Source: §5 · p. 15

Threats to validity

  • SALT-KG represents one SAP S/4HANA instance, and public datasets lack comparable expert-curated determination graphs.The generalization study tests the cascade principle rather than the full KG-guided workflow; the LLM sensitivity check covers two fields, and the temporal split coincides with COVID-19.Source: §5 · p. 15

Limitations

  • The study has not tested classical rule-mining baselines, limiting separation of KG-guided key selection from symbolic rule extraction generally.Other stated limitations include dependence on configured determinism and coarse defaults for cold-start entities.Source: §5 · p. 16

Future directions

  • Future work proposes temporal-windowed cascades for threshold-based systems and hybrid cascade-GNN models for unseen or recently reorganized entities.The hybrid approach would use cascade predictions as features or priors for GNN training.Source: §5 · p. 16

6 Conclusion

The paper concludes that KG-guided cascade lookups outperform evaluated GNN baselines across the eight enterprise fields, but their advantage depends on recoverable rules. On public relational tasks, cascades match or trail GNNs when temporal or neighborhood reasoning matters.

Conclusion

  • KG-grounded cascade lookups outperform evaluated GNN baselines on all eight SALT-KG enterprise fields.The workflow uses LLM-guided SQL profiling and deploys the resulting lookups as lightweight JSON files.Source: §6 · p. 16

AI Usage

  • The author reports using LLMs in offline rule discovery and for manuscript readability and grammar assistance, with all content reviewed by the author.Source: AI Usage · p. 17

Disclosure of Interests

Related papers

Loading 2610.34258v1…