Benchmarking Streaming Evolutionary Ensemble Learning under Shifting Imbalanced Data

Streaming classification tasks with shifting imbalanced data distributions imply that the process creating the data are non-stationary. Such a streaming context is particularly challenging because model building, at any point in time, can only be performed relative to a small (incomplete) sample of the data. We demonstrate that both crossover and ensemble learning are particularly important for competitive performance under two challenging benchmarks. One benchmark represents a Botnet detection task and the second reformulates the Forest cover type classification task for shift (as opposed to drift). Comparison with the contemporary Learn++.NSE streaming classifier indicates specific advantages and disadvantages as viewed from the perspective of null-bias and prequential performance metrics.