An Adaptive Ensemble Classifier for Handling Recurring Concepts
Tinofirei Museba, Fulufhelo V. Nelwamodo, Khmaies Ouahada · 2019
The assumption with many learning algorithms is that the underlying distribution of the data is static. However, for many real world applications, data is streaming and collected over an extended period of time. Learning in such dynamic and nonstationary environments presents a challenge not common in static domains as the statistical properties of the target variable which the model is trying to predict change over time, a phenomenon known as concept drift. The presence of concept drift can potentially cause a significant accuracy deterioration of an exploiting classifier. Furthermore, previously learnt concepts may reappear and reusing previously learnt models can optimize the learning process in terms of predictive accuracy and processing time. In this paper, we propose handling recurring concepts in time evolving environments with the Diversity Based Ensemble for handling recurring concepts (DERC), a learning algorithm that preserves previously learned diverse models and trains every model preserved with the new data. Empirical studies on one synthetic data set and one real world data set, all associated with concept drift demonstrate that DERC can effectively handle recurring concepts than other two state of the art approaches.