DynED: Dynamic Ensemble Diversification in Data Stream Classification

What is it about?

Ensemble methods are a popular choice for classification tasks because they perform well. However, it's tough to maintain high accuracy when data changes over time, a phenomenon known as concept drift. It's been observed that having a diverse set of components in an ensemble can help improve accuracy in such situations. But not all components contribute equally to the performance, so it's important to pick the ones that are both diverse and perform well. We propose a new method (DynED) for building and maintaining ensembles that takes into account both the diversity and accuracy of components. This method uses something called Maximal Marginal Relevance (MMR) to dynamically select the best components while building the ensemble. Our experiments on various real and synthetic datasets show that our method outperforms five other well-known methods in terms of average accuracy.

Featured Image