Predicting concept drift via dynamic Naïve Bayes

Abstract

Listen

Translate article

Concept drifts occur when the relationship between streaming input and target response changes. Consequently, prediction models need to be updated to maintain prediction accuracy. This paper explores the idea of extending the Naive Bayes method to predict the time a predictive model must be changed due to the concept drift. Our dynamic Naive Bayes method maintains a simple data structure to efficiently predict when to update a predictive model to reflect the substantial changes in data distribution. Simulation results show that our method can significantly reduce the frequency of model updates while maintaining high prediction quality when compared to the traditional approaches where predictive models must be rebuilt periodically.