Top-Weighted Structural Analysis of Ordered List Averaging for Medical Knowledge Extraction
Ricardo Sánchez-de-Madariaga, Mario Pascual Carrasco, Adolfo Muñoz Carrero · Mathematics · 2026
Ordered List Averaging (OLA) was previously introduced as a methodology for evaluating machine learning algorithms according to their ability to prioritize medically relevant feature subsets. The original methodology demonstrated, through Friedman–Nemenyi statistical analysis applied to probability distributions of related datasets, that OLA can identify statistically supported machine learning algorithms for medical knowledge extraction. However, the internal structural organization of OLA-generated rankings has not previously been investigated. This work presents a structural analysis framework for OLA based on top-weighted ranking-agreement measures. Rather than proposing a new machine learning model-selection algorithm, the objective is to determine whether the ordering induced by OLA scores exhibits an intrinsic structural organization that provides complementary evidence supporting the interpretation of OLA rankings. To this end, ranking signatures generated by OLA are compared using Kendall distance, while the agreement between score-based and structure-based rankings is evaluated through Top-k overlap and Rank-Biased Overlap (RBO). Experiments conducted on nine biomedical datasets show that datasets previously exhibiting statistically validated OLA behavior consistently present stronger top-weighted structural agreement than datasets with weaker statistical support. Furthermore, the observed structural patterns remain stable across different values of the RBO persistence parameter and are robust to variations in dataset size. The results indicate that OLA rankings possess a meaningful and reproducible structural organization beyond their numerical scores. Consequently, top-weighted structural agreement provides complementary evidence supporting the interpretation of OLA rankings and strengthens the practical interpretation of OLA-based algorithm selection, particularly in situations where only a single working dataset is available and statistical validation through Friedman–Nemenyi cannot be performed.