A Novel Deep Ensemble Learning Framework for Classifying Imbalanced Data Stream

Abstract

Machine learning has emerged as one of the most indispensable fields of this era in mining stream data. The stream data have the tendency to change their characteristics over time. Mining imbalanced stream is a research demanding subfield of this area. In imbalanced data, one of the target classes has much less instances than other class. The imbalanced data may differ either in their ratio between majority and minority class or dimension or the number of classes. The performance of the classifier is affected by variety of imbalanced data set used for training the classifier. Because, the learning of classifier is different for different types of data sets. Imbalanced data can result due to rare events which can have negative impact on society. Most traditional data mining algorithm misclassifies the minority class of the imbalance data sets or considers it as noise. Therefore, the decision is biased toward majority class and hence reduces the accuracy and overall performance of the algorithm. The algorithms for classifying imbalanced data sets thus demand high adaptability to changes in the majority and minority class ratios. The performance of traditional machine learning algorithms is enhanced by applying ensemble method and deep learning approaches. Purpose: This paper proposes a framework based on deep ensemble learning for classifying imbalanced stream of data. Methodology: The deep ensemble methods ensemble multiple base learners. While deep learning approach is applied to improve the performance by extracting lower-level features and feeding them forward for the next layer to identify higher level features. Results: In this method, the effect of highly imbalanced classes is reduced by uniting the ensemble method with deep learning. The accuracy of the classifier is improved in terms of not only accuracy but also categorical accuracy. Conclusion: In addition to accuracy, other performance measures like categorical prediction accuracy, training accuracy and prediction accuracy have also been compared which are crucial metrics for evaluating imbalance stream of data.