KGEDO: A Knowledge-Guided Explainable Decision Optimisation Framework for Intelligent Accounting Information Systems

Abstract

Artificial intelligence has enabled accounting information systems (AIS) to advance beyond transaction processing to predictive risk analytics. However, high predictive accuracy alone is not sufficient for management use: black-box scores do not explain how a decision threshold represents asymmetric business costs, which accounting relationships create risk, or what workable adjustments could change a negative rating. This study presents KGEDO, a knowledge-guided explainable decision optimization framework that combines constrained what-if simulation, cost-sensitive prediction, accounting-domain encoding, and local explanation. Robust profitability, liquidity, solvency, leverage, exposure, and control-risk group scores, with rule violation indicators, encode accounting knowledge; a utility layer selects decision thresholds under a five-to-one false-negative cost assumption; an additive explanation engine generates auditable reason codes; and a gradient-boosted predictor estimates risk. The evaluation used three public benchmarks, Taiwanese and Polish corporate bankruptcy, and external audit fraud risk, with five-fold stratified cross-validation. KGEDO obtained mean ROC-AUC values of 0.939, 0.967, and 0.999, respectively. Averaged across tasks, its PR-AUC nearly matches raw XGBoost, yet the predicted decision cost drops by 11.2%. Constrained scenarios reversed 85–100% of the 40 highest-risk cases in each dataset. Taken together, the results argue for causal validation and manager-centered review ahead of any operational rollout: KGEDO reads less as a marginally better classifier and more as a decision-support architecture in its own right.