Popov I.E. Interpretation of Mathematical Models Based on Microwave Radiothermometry Data
- Details
- Hits: 8
https://doi.org/10.15688/mpcm.jvolsu.2025.2.6
Illarion E. Popov
Postgraduate Student, Department of Mathematical Analysis and Function Theory, Volgograd State University
This email address is being protected from spambots. You need JavaScript enabled to view it.
,
https://orcid.org/0000-0002-0997-8721
Prosp. Universitetsky, 100, 400062 Volgograd, Russian Federation
Abstract. The article considers the problem of increasing the interpretability of solutions of mathematical models while maintaining high prediction accuracy. The main attention is paid to the integration of highly accurate neural networks with an ensemble of interpretable models whose mechanisms are transparent and amenable to analytical description. A method for forming a justification based on ensemble decisions consistent with the prediction of the neural network is proposed. The agreement is achieved by comparing the confidence levels of the models after preliminary calibration, the need for which is due to the overconfidence effect characteristic of some machine learning models. A method has been developed for selecting interpretable models of an ensemble of classifiers whose estimates on a specific object are as close as possible in terms of confidence to the output of the neural network. This allows forming justifications containing both arguments in favor of the decision made and possible alternative opinions. To increase the flexibility of interpretation, an extended definition of a highly informative feature has been introduced, including categorization of values by the degree of their characteristic for different classes. It is shown that the transition from binary to categorical features contributes to the growth of classification accuracy and increases its overall efficiency. Additionally, a method for constructing informative intervals of features has been developed, which allows increasing their informativeness – separating ability. Based on the obtained intervals, algorithms for refining semi-definite labels and correcting the training sample in order to improve its quality and representativeness have been proposed. The proposed approaches have been tested on the problem of breast cancer diagnostics using microwave radiothermometry data. The results of computational experiments confirm that the use of categorical interpretable features in combination with model calibration allows for a significant increase in classification accuracy and the validity of decisions made.
Key words: parallel computing, gravitational systems, OpenMP, processor architecture, Hyper-Threading technology.
Interpretation of Mathematical Models Based on Microwave Radiothermometry Data by Popov I.E. is licensed under a Creative Commons Attribution 4.0 International License.
Citation in English: Mathematical Physics and Computer Simulation. Vol. 28 No. 2 2025, pp. 62-80
 popov.pdfURL: https://mp.jvolsu.com/index.php/en/component/attachments/download/1235 | 2 Downloads |