Performance Degradation of Multi-class Classification Model Due to Continuous Evolving Data Streams

Abstract

Online machine learning plays a pivotal role in the 4th Industrial Revolution (IR 4.0). IR 4.0 requires real-time data analysis (classification or prediction) using the streaming data. However, mostly the data streams contain nonstationary (variability features) characteristics, such as concept drift and class imbalance. The issues of concept drift and class imbalance adversely affects the accuracy of the classification models. The classification accuracy is even more affected when these issues arrived at the same time (joint problem). Some efforts have been made in the literature to cope with the joint problem of class imbalance and concept drift in online learning, but the existing solutions are limited to binary class classification and these solutions do not work for multi-class classification. Besides, the literature doesn’t mention the exact correlation between the critical factors of concept drift and class imbalance. Also, the tuning parameters of the multi-class classification models, which can help in improving the classification accuracy, are unknown. However, to resolve the joint problem of concept drift and class imbalance in online multi-class classification models, it is essential to determine the exact correlation and find the tuning parameters, which could be helpful to provide a more dynamic approach for avoiding the performance degradation of the multi-class classification models. Therefore, to resolve this issue, this study aims to determine the correlation between the concept drift and class imbalance, identify the tuning parameters of multi-class classification models and propose a dynamic solution based on these findings. Our proposed dynamic approach could be effectively utilized in various online machine learning based (real-time) streams analysis, which is desirable for IR 4.0.