Naučno-tehničeskij Vestnik Informacionnyh Tehnologij, Mehaniki i Optiki (Jun 2024)
Method for generating information sequence segments using the quality functional of processing models
Abstract
The constantly emerging need to increase the efficiency of solving classification problems and predicting the behavior of objects under observation necessitates improving data processing methods. This article proposes a method for improving the quality indicators of machine learning models in regression and forecasting problems. The proposed processing of information sequences involves the use of input data segmentation. As a result of data division, segments with different properties of observation objects are formed. The novelty of the method lies in dividing the sequence into segments using the quality functional of processing models on data subsamples. This allows you to apply the best quality models on various data segments. The segments obtained in this way are separate subsamples to which the best quality models and machine learning algorithms are assigned. To assess the quality of the proposed solution, an experiment was performed using model data and multiple regression. The obtained values of the quality indicator RMSE for various algorithms on an experimental sample and with a different number of segments demonstrated an increase in the quality indicators of individual algorithms with an increase in the number of segments. The proposed method can improve RMSE performance by an average of 7 % by segmenting and assigning models that have the best performance in individual segments. The results obtained can be additionally used in the development of models and data processing methods. The proposed solution is aimed at further improving and expanding ensemble methods. The formation of multi-level model structures that process, analyze incoming information flows and assign the most suitable model for solving the current problem makes it possible to reduce the complexity and resource intensity of classical ensemble methods. The impact of the overfitting problem is reduced, the dependence of processing results on the basic models is reduced, the efficiency of setting up basic algorithms in the event of transformation of data properties is increased, and the interpretability of the results is improved.
Keywords