Novel Technique(s) for Concept Drift Detection and Handling

Abstract

In today’s world, learning in the presence of dynamic environments, where continous change and development are evident, is a challenging task. Recent advances in technology have witnessed an increase in the number of real world applications which include spam filtering, fraud detection, weather forecasting, sensors, smart cities, health monitoring etc. Data generated from such sources in form of streams tends to evolve with the course of time. The predictive models which are trained using such data tend to become obselete with time, resulting into poor adaptabilty to the underlying drifting distributions. In terms of machine learning and data mining, the change in the statistical properties of data is known as concept drift. Such changes causes degradation in the performance of the learning systems since the models that were built on the old data are no longer consistent with the new data. To address the problem of concept drift, efficient learning models which can monitor the evolving distributions and update themselves regularly are required. These models should detect the drifts and handle them timely by using adaptive learning techniques, to overcome the deteriorating performance. Various learning methods which include single learners as well as ensemble based modelling which utilize drift detectors, are used in literature to handle evolving data streams. This thesis proposes three techniques for concept drift detection and handling. First one, a hybrid diversity based ensemble approach, called Ensemble Based Online Diversified Drift Detection (En-ODDD), combines explicit drift detection and adaptive techniques deal with drifting distributions. In second approach, Two-Level Pruning based Ensemble with Abstained Learners (TLP-EnAbLe), similarity based pruning strategy has been proposed for adapting to all types of drift patterns. The third approach, Dynamically Adaptive and Diverse Dual Ensemble (DA-DDE) utilizes the characteristics of both online and block-based ensemble techniques for concept drift handling. It proposes a dual ensemble mechanism for separetely handling abrupt and gradual drifts. It is based on usage of novel Dynamic Dual Selective Voting Mechanism (DDSVM) for ensemble selection and hypothesis generation. Performance of the proposed approaches has been evaluated by conducting comparative analysis with existing concept drift techniques and through the standard evaluation parameters which include classification accuracy, kappa statistic, train time, test time, memory consumption etc. Experiments conducted using several real datasets and artificially generated streams of data, with variety of drift patterns, indicate that all the three approaches handle the concept drift scenarios effectively giving better classification results.