Semantic Data Fragmentation for Identification of Covariant Conceptual Drift in Machine Learning Models

Abstract

Such types of data drift in ML models as actual positive and negative drifts, as well as varieties of fragmentary drifts in different directions, are considered. An overview of the current state of the problem is given, which highlights the methods of sliding window, trigger ensemble of models, covariant shift, personalization drift correction, season correction, online learning method, low precision sampling, monitoring and clipping features. A modernized theory of binary relations was used to design the new method. As an example, the subject area "communication services" is considered, for which a special architecture of the ontological knowledge model is designed. The new drift correction method is the basis of a new technology for designing classification, regression and forecasting models for specifically formalized subject areas. When choosing the scope of the "sliding window", the structure of the knowledge model is taken into account first of all. The input features of the training data set are grouped according to the structure of the concepts and relationships of the knowledge base. The resulting data drift compensation technology makes it possible to improve the ROC AUC accuracy characteristics of ML models from 0.74 to 0.80, which makes it possible to evaluate the technology as an effective means of automatic correction. The new drift correction method is the basis of a new technology for designing classification, regression and forecasting models for specifically formalized subject areas. When choosing the scope of the "sliding window", the structure of the knowledge model is taken into account first of all. The input features of the training data set are grouped according to the structure of the concepts and relationships of the knowledge base.