Attention-Guided Efficient Structure Reassembly Through Reverse Contribution-Driven Feature Distillation and Inconsistency-Resilient Sample Filtering

Excessive features and/or inconsistent samples in practical complex system modeling often misguide decision-makers’ attention. To address this challenge, a novel attention-guided efficient structure reassembly approach is proposed through reverse contribution-driven feature distillation and inconsistency-resilient sample filtering. Specifically, feature distillation is achieved through attention-guided reverse identification of original key features via traditional dimensionality reduction techniques with added feedback-based contribution analysis. Sample filtering is realized by detecting inconsistent samples using artificial errors-based clustering results. These two mechanisms also form the main theoretical contributions of this work, which place the attention of complex systems modeling back to only the key features and samples, thus saving limited resources, improving efficiency, and maintaining high accuracy. Validation across three case studies demonstrates that: (1) feature distillation can reduce an average of 39.34% of original features but only results in 1.82% classification accuracy reduction across twelve UCI benchmarks; (2) sample filtering for pipeline leakage detection filters 0.79% inconsistent samples, leading to more than 10% improvement in modeling accuracy on a pipeline leakage detection task using BPNN, RBF, and GPR models; (3) reassembling an efficient structure with only 53.86% of the original features reduces the building tilt rate (BTR) prediction error by 13.23% (from MAE = 0.1360 to MAE = 0.1180); and (4) the proposed approach demonstrates strong model-agnostics, across different baseline models.