Do fenômeno à reação: descritores da mudança de conceito como evidências para estratégias de reação

Abstract

The rapid evolution of data streams poses significant challenges to machine learning models, which are frequently exposed to the phenomenon of concept drift, characterized by changes in the joint distribution of instances over time. Although adaptive approaches such as incremental learning and full retraining are widely employed, the literature still presents a gap regarding the use of concept drift descriptors as evidence to guide the choice of the most suitable reaction strategy. This thesis aims to demonstrate that concept drift descriptors, including severity, influence zone, speed, frequency, recurrence, and predictability, can support the definition of a reaction strategy capable of preserving classifier performance in dynamic scenarios. To this end, three stages of investigation were conducted: (i) a systematic review and the proposal of a new taxonomy for unsupervised concept drift detection methods; (ii) an experimental analysis of the impact of descriptors on classifier performance across 32 distinct scenarios and 5 synthetic datasets; and (iii) a comprehensive evaluation of five reaction strategies to concept drift (static, full retraining, incremental learning, forgetting, and oracle) under different descriptor configurations, complemented by statistical tests, effect size analysis, and an ablation study. The results show that the adequacy of the reaction strategy is directly related to the characteristics of the drift. It was found that descriptors such as severity and speed strongly influence classifier performance, while recurrence and frequency are determinant in scenarios of pattern reappearance. Furthermore, the widely adopted strategy in the literature, full retraining, proved not to be fully effective, as in certain situations approaches such as incremental learning with forgetting, or even no reaction, yielded superior performance. The main contributions of this research include: (i) the proposition of a systematic taxonomy for unsupervised concept drift detection methods; (ii) the empirical characterization of the interrelationships among descriptors and their impact on classifier performance; (iii) the evidence-based definition of reaction strategies to concept drift; and (iv) the presentation of a robust and replicable experimental framework. The findings highlight the importance of considering the nature of concept drift when developing adaptive learning systems, offering support both for academic research and for practical applications in sensitive domains such as healthcare, security, and finance.