Dynamic Indicator Cluster Extraction for Combat Systems of Systems Based on Spatiotemporal Graph Association Learning

Abstract

To address indicator associations that vary with mission phase, temporal evolution, and task conditions in highly dynamic air and missile defense missions, this paper proposes STG-ICENet, a method for extracting dynamic indicator clusters. The method jointly models indicator time windows, lagged local relation graphs, and a global graph derived from training scenarios. It obtains spatiotemporal representations through graph convolution, predicts candidate relations with each query indicator as the center, and combines an adaptive threshold with Top-K selection to output dynamic indicator clusters. To prevent information leakage, reference relations for the current window are used only for supervision and evaluation. On the 7,680 test queries reported in the available material, Precision@10, Recall@10, and F1@10 are all 0.5466; NDCG@10 is 0.6509; MRR is 0.9734; ROC-AUC is 0.8178; and PR-AUC is 0.5976. The results show that the framework balances first-hit performance, ranking quality, and relation discrimination, and provides traceable indicator subsets for phase-specific effectiveness analysis. Cross-scenario generalization, ablation studies, and computational cost remain to be validated.