Emergency decision-making method for carrier-based aircraft support operations based on multi-view graph contrastive learning

To address the issues of traditional emergency decision-making methods for carrier-based aircraft support operations, such as excessive reliance on historical case databases and commander experience, significant data sparsity, and insufficient mining of correlations between emergency features and sub-plans, this paper proposes a multi-view graph contrastive learning(MVGCL)model.

This study formulates the emergency decision-making problem for carrier-based aircraft support operations as a task of recommending multiple sub-plans based on the multi-attribute characteristics of emergencies. First, the model constructs two types of auxiliary views: non-augmented and augmented. On the one hand, it builds an emergency feature collaboration graph and a sub-plan collaboration graph based on the intrinsic relationships between emergency features and sub-plans. On the other hand, it generates two augmented auxiliary views by applying approximate singular value decomposition (ApproxSVD) and noise masking to the emergency feature-sub-plan interaction graph. By fully leveraging the self-supervised information from the interaction graph and multi-view contrastive learning, the model improves recommendation accuracy. Additionally, a debiased contrastive loss function is adopted to mitigate false-negative interference in traditional contrastive loss by correcting negative sample sampling bias, enhancing both the accuracy and robustness of recommendations.

The experimental results on the simulated dataset demonstrate that the proposed model outperforms both single-view graph contrastive learning models and traditional case-based reasoning methods in terms of recall, precision, normalized discounted cumulative gain (NDCG), and mean reciprocal rank (MRR). Specifically, when the number of recommended sub-plans is 10, the Recall and NDCG of the proposed model show improvements of 1.14% and 2.39%, respectively, compared to the best graph contrastive learning baseline HGCL, and significant enhancements of 64.36% and 43.1%, respectively, over the case-based reasoning method CBR.

The experimental results validate the effectiveness of the proposed multi-view graph contrastive learning framework in modeling self-supervised information and alleviating the sparsity of interaction data.