Adaptive Segmentation of Information Sequences for Machine Learning Modular Regression Models

The research objective is to construct an adaptive model for modular machine learning structures that improves the processing quality of information sequences. The novelty of the proposed methodology is that it can identify segments of an information sequence obtained using various methods and assign models with the best quality indicator values to subsequences. Classical methods allow tuning of the model to the entire data sample. The improvement consists of the proposed solutions that consider the inverse problem of forming segments of data sequences, such that their properties correspond to the processing model. The proposed methodology was tested on various models and datasets. Segmentation and assignment of regression models with the best characteristics to individual segments allow the reduction of the mean square error (MSE) and mean absolute error (MAE) to 8%. The findings show an opportunity to increase of 5-8% for weak LR, SVM, and GR models, while strong DT, CNN, ANN, ANFIS, and XGBoost models improve by 1-4% in scenarios with limited data. Segmentation enables more efficient training and reduces sensitivity to noise and outliers. The proposed solution allows the selection of segmentation strategies and model combinations considering local data properties. Its application enables the implementation of existing machine learning architectures to improve the quality of training and analysis of information sequences and increase adaptability, scalability, and interpretability.