Comment
Author: Admin | 2025-04-28
And is considered one of the state-of-the-art concept drift detectors. A common problem with sliding windows is that the width usually needs to be predefined. Varying the width of the window has an impact on performance, thus applying the correct width value is important. Typically this is done by means of some user-defined parameter. ADWIN, however, is parameterless and the sliding window is sized dynamically by the algorithm itself. It also provides excellent performance due to the use of buckets and the adaptation of the exponential histogram to for compression.4.3 Block-based ensemble detectorsIn machine learning, an ensemble refers to a group or collection of classifiers that work together to achieve greater predictive performance. Block-based ensembles process data in blocks, or chunks, of some specified size. The performance of block-based ensemble methods is based heavily on the chunk size. Similarly to that of sliding windows, larger chunks tend to produce more accurate classifiers but may contain concept drift within themselves. Alternatively, smaller based chunk sizes are typically more effective at drift detection but produce inferior performing classifiers.The Streaming Ensemble Algorithm (SEA) was first proposed by Street and Kim [65] and is a block-based ensemble learning algorithm. Individual classifiers are constructed from examples read in sequential blocks (chunks), which are then added to a fixed size ensemble. If the ensemble is full, then the worst performing classifier is removed from the ensemble entirely.In their experiments, C4.5 [61] classifiers are used for building the ensemble. The output prediction is given as the simple majority voting of the entire ensemble. Results for testing with concept drift showed that the algorithm was capable of recovering quickly by discarding classifiers trained on the outdated data.A notable drawback to SEA is the way in which classifiers are replaced. Merely replacing the worst performing classifier with the most recently trained has the potential to still leave several, outdated and poorly performing classifiers in the ensemble, depending on the predetermined ensemble size.This is improved upon by Wang et al. [68]’s Accuracy Weighted Ensemble (AWE), a block-based algorithm which trains a new learning model with each new chunk of data in a similar fashion to that of SEA. Where AWE improves upon SEA is in the model replacement. AWE uses a version of the mean square error to select n best classifiers to construct an entirely new ensemble, thus removing all other outdated and poorly performing classifiers. The algorithm for
Add Comment