Concept Drift and Model Decay Detection using Machine Learning Algorithm
Pragati Aravind Nayak, Pavithra Sriganesh, K.M Rakshitha, Manoj Kumar M V, Prashanth B S, H R Sneha · 2021
Machine learning mainly focuses on building applications that can automatically learn periodically and improve their working experience without being explicitly programmed again. Data extracted from dynamic distributions for real-world applications lead to a phenomenon called concept drift. This hitch could encounter when the dependency and relationship of the input data features and class variable varies gradually. So, the model must be capable of processing this data and swiftly adapt to these changes. In this context, we opt for an unsupervised approach, which uses a sliding window mechanism to detect drift in the concept by analyzing the variations in the distribution of the new samples. It uses a classifier that does not have any inbuilt drift detection and adoption mechanism. Experimentation is done on various drift detectors using several data sets. The performance of our method comparatively is higher than D3, Adwin, Page-Hinkley, and OCDD methods.