A partial labeling framework for multi-class imbalanced streaming data
Elaheh Arabmakki, Mehmed M. Kantardzic, Tegjyot Singh Sethi · 2017
Imbalanced data streams are found in many real world applications such as spam email detection, and internet traffic data. The classification of such data is challenging, since data stream usually changes, and the model should be updated to maintain the performance. However, obtaining the true labels of the samples to build a new model is not easy, since labeling is expensive and time consuming. Additionally, existence of the multiple and imbalanced classes may cause to lose performance over one class while trying to gain on another. In this paper, we propose RLS-Multi (Reduced Labeled Samples-Multiple class) which is a classification framework for the multi-class and evolving imbalanced data stream. RLS-Multi handles the data with multiple classes, and it uses a small fraction of the data to update the model. RLS-Multi is compared with McELM, and VWOS-ELM which are two fully labeling approaches for classification of the imbalanced and multi-class data stream. The experimental results show that the performance of the RLS-Multi is not significantly different from the two other techniques, requiring only up to 25% of the samples to label for majority of the data sets, on average.