Identifying dependencies in mathematical data via predictive importanceNovoselec, Matej (Avtor) Todorovski, Ljupčo (Mentor) Narvaez Denis, David Eliecer (Komentor) AI for mathematicsmachine learningpredictive importancedependency detectiononline encyclopedia of integer sequences (OEIS)cubic vertex-transitive graphs (CVTG)We consider the problem of identifying subsets of variables that are likely to participate in an underlying (possibly unknown) relationship within a mathematical database. By treating each variable as a target, the proposed approach aggregates various predictive importance measures from a diverse set of machine learning algorithms to construct a weighted network of pairwise dependencies. We hypothesize that groups of variables with high mutual predictive importance correspond to candidates for meaningful mathematical relationships. This hypothesis is tested on a curated subset of integer sequences from the OEIS with known interdependencies, as well as a census of cubic vertex-transitive graphs. Our results on the OEIS experiment show that when algorithms are trained on sufficiently long prefixes of the sequences, the method identifies variables involved in known relationships and suggests candidate relationships, achieving an Area Under the Precision-Recall Curve (PR-AUC) more than four times that of a random baseline. The application to the graph census confirms the framework's ability to generalize across different mathematical domains, even in the presence of challenging data constraints.20262026-09-10 08:15:11Magistrsko delo/naloga187295sl