Cross-testing methodology for pattern learning and model transfer in rare fog events detection
Gaetano Zazzaro · Pattern Recognition · 2025
In this paper, a novel mathematical relation based on the k -Nearest Neighbor algorithm is introduced, designed to compare numerical datasets and enable their augmentation with additional examples to train more effective models. This relation forms the foundation of the proposed methodology, termed Cross-Testing . This comparison strategy effectively addresses the class imbalance problem in Pattern Recognition by enabling the transfer of knowledge across datasets connected by the relation and oversampling the minority class with real data. It is also verified that merging related datasets enhances model performance compared to training on each dataset separately. To validate the methodology, a meteorological case study is conducted, identifying similar weather conditions across 25 Italian airport stations and detecting rare fog events, which are challenging due to class imbalance. This case study demonstrates the efficacy of the methodology as a valuable tool for Pattern Recognition and Transfer Learning.