THE DYNAMIC ENSEMBLE GENERATION AND CLASSIFICATION ALGORITHM FOR MINING HIGH SPEED DATA STREAMS

The class evolution, the occurrence of class appearance and disappearance, is an important research topic for mining high speed data streams. All previous studies implicitly observe class evolution as a temporary change, which is not true for many real-world problems. This paper concerns the situation where classes emerge or disappear gradually. A class-based ensemble method viz. The CBCE (Class-Based ensemble for Class-Evolution) is projected by keeping a base learner for every class and vigorously updating the base learners with new data, it can quickly adjust to class development. A novel under-sampling method for the base learners is also intended to handle the dynamic class-unbalance problem caused by the gradual evolution of classes. The data-based studies reveal the effectiveness of CBCE in a variety of class evolution state in comparison to existing class evolution adaptation methods.