Learning Concept Drift Using Adaptive Training Set Formation Strategy

Abstract

We live in a dynamic world, where changes are a part of everyday ‘s life. When there is a shift in data, the classification or prediction models need to be adaptive to the changes. In data mining the phenomenon of change in data distribution over time is known as concept drift. In this research, we propose an adaptive supervised learning with delayed labeling methodology. As a part of this methodology, we introduce an adaptive training set formation algorithm called SFDL, which is based on selective training set formation. Our proposed solution considered as the first systematic training set formation approach that take into account delayed labeling problem. It can be used with any base classifier without the need to change the implementation or setting of this classifier. We test our algorithm implementation using synthetic and real dataset from various domains which might have different drift types (sudden, gradual, incremental recurrences) with different speed of change. The experimental results confirm improvement in classification accuracy as compared to ordinary classifier for all drift types. Our approach is able to increase the classifications accuracy with 20% in average and 56% in the best cases of our experimentations and it has not been worse than the ordinary classifiers in any case. Finally a comparison study with other four related methods to deal with changing in user interest over time and handle recurrence drift is performed. Results indicate the effectiveness of the proposed method over other methods in terms of classification accuracy.