2026 , Volume 31, № 1, p.92-105

Purpose. The paper addresses the problem of weakly supervised multiple instance learning, where sets of objects referred to as “bags” are analyzed. Each object is represented by a set of observations of certain features. A binary classification case is considered: one class is conventionally labelled as positive, and the other as negative. A bag is labelled as positive if it contains at least one positive object (the specific object is unknown); otherwise, the bag is labelled as negative. The goal is to predict classes for new bags to achieve the best quality metrics.

Methodology. Machine learning methods adapted to the problem are employed: 𝑘-nearest neighbours (𝑘NN), informative feature selection, sample filtering, and an ensemble approach to constructing decision functions. Additionally, the function of rival similarity (FRiS) is used to evaluate the degree of “unusualness” of bags. An experimental study and comparison with existing methods are conducted on the real-world problem of identifying proteins containing structures with a thioredoxin fold.

Findings. A method for solving the problem was developed, utilizing informative feature space selection, filtering (self-correction) of the training set, and the voting on a set of decision functions. The results of solving the protein identification problem were compared with a number of well-known algorithms using quality metrics.

Originality/value. The developed method enables the selection of the most informative feature sets, which is crucial for improving the quality and interpretability of solutions, as well as selfcorrection of the training set, reducing the impact of various errors associated with inaccurate labelling, outliers, etc. In numerical experiments with protein structure recognition data, comparison results with a number of well-known algorithms confirmed sufficiently high efficiency of the proposed method according to the balanced accuracy metric.