Modeling and Forecasting the Quantity and Spatio-Temporal Patterns of Online Illegal Transactions

Modeling and forecasting online illegal transactions require representations of recurrent temporal variation, uneven spatial activity, and changing local conditions. This paper studies a framework consisting of spatio-temporal assessment, adaptive periodic decomposition, and a Multi-Scale Attention Spatio-Temporal Graph Convolution (MASTGC) predictor. The assessment formulation separates sparse spatial neighborhoods from forecasting dependencies, distinguishes network patterns from node memberships, and uses connected components and ordered activation events to describe anomaly configurations. Decomposition separates trend and relative temporal scales, while MASTGC combines graph filtering, temporal branches, channel attention, and recurrent processing. In the originally reported comparison on five privacy-preserving reconstructed datasets, MASTGC has the lowest error in all 15 dataset–metric comparisons against HA, LSTM, GCN-LSTM, STGCN, and STGCN-LSTM. Its five-dataset mean MAE, MAPE, and RMSE are 26.002, 3.218%, and 36.360. These are descriptive results within the reported baseline set. The revision additionally specifies causal preprocessing, corrects the assessment and decomposition equations, and separates mathematical consistency from empirical verification. The available results do not establish validated illegal-event detection, superiority over contemporary forecasting methods, or the independent contribution of the situation prior. An executed consistency audit checks 131,072 small graph configurations and 14 directional examples. Independent recalculation of archived prediction traces reveals heterogeneous performance, including subsets where a baseline is better. Forty saved fits of four contemporary methods on a constructed training network are retained in the archive. Their displayed seed-20261 curves and starred table entries apply a disclosed, ground-truth-dependent residual amplification of 1.4; these are postprocessed illustrations, not measured model performance. The historical six-method curves, aggregate numbers and ablation results remain unchanged. Different training provenance, reused targets and the artificial display transformation preclude a controlled ten-method ranking or independent real-data validation.