Critical risk node identification in enterprise resource planning heterogeneous knowledge graphs

Enterprise Resource Planning (ERP) systems record high-volume, multi-typed transactions that are difficult to audit with manually specified rules alone. Existing graph-based anomaly detection methods often simplify ERP data into homogeneous graphs or use attention mechanisms that do not explicitly model the temporal ordering of audit events. We propose AuditGain, a framework for constructing ERP heterogeneous knowledge graphs and identifying critical risk nodes with an Adaptive Graph Attention Network. The model combines three main components: type-specific feature extraction for ERP entities, LSTM-enhanced multi-head attention for temporally ordered neighborhoods, and a Diffeomorphic Graph-Adaptive Activation Function (DiGRAF) for graph-conditioned nonlinear transformations. A community-level memory module based on training-period spectral clustering is used to retain cross-period context. Experiments on three proprietary ERP audit datasets and two public fraud detection benchmarks (YelpChi and Amazon) show consistent gains over heterogeneous, temporal, and fraud-oriented graph baselines. On ERP-Audit-A, AuditGain outperforms TGN by 4.2 percentage points and the strongest recent temporal baseline in our comparison by 3.5 percentage points in F1-score. Because the ERP datasets cannot be released under non-disclosure agreements, we provide public benchmark experiments, configuration files, and a synthetic ERP graph generator to support independent checks of the evaluation pipeline. The source code is publicly available at https://github.com/Drake-crz/AuditGain .