An ELM-based model with sparse-weighting strategy for sequential data imbalance problem

In many practical engineering applications, online sequential data imbalance problems are universally found. Many traditional machine learning methods are hard to improve the classification accuracy effectively while solving these problems. To get fast and efficient classification, a new online sequential extreme learning machine algorithm with sparse-weighting strategy is proposed to increase the accuracy of minority class while reducing the accuracy loss of majority class as much as possible. The main idea is integrating a new sparse-weighting strategy into the present data-based strategy for sequential data imbalance problem. In offline stage, a two phase balanced strategies is introduced to obtain the valuable virtual sample set. In online stage, a dynamic weighting strategy is proposed to assign the corresponding weight for each sequential sample by means of the change of sensitivity and specificity in order to maintain the optimal network structure. Experimental results on two kinds of imbalanced datasets, UCI datasets and the real-world air pollutant forecasting dataset, show that the proposed method has higher prediction accuracy and better numerical stability compared with ELM, OS-ELM, meta-cognitive OS-ELM and weighted OS-ELM.