Detecting data drifts in predictive models by explainable artificial intelligence tools

Concept drift refers to changes in data sets over time that negatively impact theperformance of predictive models. This thesis addresses the limitations of traditionalstatistical methods used to detect concept drift and proposes a new method based onPartial Dependence Profile (PDP), an eXplainable Artificial Intelligence (XAI)technique: Profile Drift Detection (PDD). This new method evaluates changes in PDPcurves using metrics such as Euclidean distances between profiles, Euclidean distancesbetween their derivatives, and the Profile Disparity Index (PDI).The study compares traditional statistical tests and the Profile Drift Detection(PDD) method using synthetic and real-world data sets commonly used in the literature.The data sets are divided into different batch sizes for analysis. PDD generally detectsfewer instances of concept drift while achieving similar or better model performancecompared to statistical tests, and it provides consistent results when batch sizes vary.Additionally, PDD enables a more detailed examination of the relationship between themodel and the data. The performance of the proposed method is investigated throughexperimental studies conducted on synthetic and real-world data for different batch sizes.The usability of the method is discussed based on the findings.