Novel Class Detection in Concept Drifting Data Streams Using Decision Tree Leaves

Abstract

Concept drifting data streams often occurs in weather forecasting, intrusion detection and other applications. One of the difficulties with handling concept drifting data streams is the existence of novel classes in the data stream that arrives after the training of the model on the existing class instances. In this thesis, we present a novel class detection algorithm in concept based on the instance distribution in the decision tree leaves. Our proposed algorithm is easy to implement and use compared to complex ensemble based methods. We have tested the performance of our algorithm on several datasets and it shows significantly improved results compared to previous state-of-the-art algorithm using standard evaluation methods and metrics.