Handling Concept Drift Based on Data Similarity and Dynamic Classifier Selection

Abstract

In real-world applications, machine learning algorithms can be employed to perform spam detection, environmental monitoring, fraud detection, web click stream, among others. Most of these problems present an environment that changes over time due to the dynamic generation process of the data and/or due to streaming data. The problem involving classification tasks of continuous data streams has become one of the major challenges of the machine learning domain in the last decades because, since data is not known in advance, it must be learned as it becomes available. In addition, fast predictions about data should be performed to support often real time decisions. Currently in the literature, methods based on accuracy monitoring are commonly used to detect changes explicitly. However, these methods may become infeasible in some real-world applications especially due to two aspects: they may need human operator feedback, and may depend on a significant decrease of accuracy to be able to detect changes. In addition, most of these methods are also incremental learning-based, since they update the decision model for every incoming example. However, this may lead the system to unnecessary updates. In order to overcome these problems, in this thesis, two semi-supervised methods based on estimating and monitoring a pseudo error are proposed to detect changes explicitly. The decision model is updated only after changing detection. In the first method, the pseudo error is calculated using similarity measures by monitoring the dissimilarity between past and current data distributions. The second proposed method employs dynamic classifier selection in order to improve the pseudo error measurement. As a consequence, this second method allows classifier ensemble online self-training. The experiments conducted show that the proposed methods achieve competitive results, even when compared to fully supervised incremental learning methods. The achievement of these methods, especially the second method, is relevant since they lead change detection and reaction to be applicable in several practical problems reaching high accuracy rates, where usually is not possible to generate the true labels of the instances fully and immediately after classification.