Active broad learning with multi-objective evolution for data stream classification

Abstract In a streaming environment, the characteristics and labels of instances may change over time, forming concept drifts. Previous studies on data stream learning generally assume that true label each instance is available or easily obtained, which impractical in many real-world applications due to expensive time labor costs for labeling. To address issue, an active broad based multi-objective evolutionary optimization presented classify non-stationary stream. The newly arrived at step stored chunk turn. Once full, its distribution compared with previous ones by fast local drift detection seek potential drift. Taking diversity their relevance new into account, algorithm introduced find most valuable candidate instances. Among them, representative are randomly selected query ground-truth labels, then update model adaption. More especially, number determined stability adjacent historical chunks. Experimental results 7 synthetic 5 datasets show proposed method outperforms five state-of-the-art classification accuracy labeling cost regions accurately identified budget adaptively adjusted.