Page 20 - Data Science Algorithms in a Week

Unsupervised Ensemble Learning 5

Table 1. Traditional and Modern algorithms

CLUSTERING ALGORITHMS BASED ON ENSEMBLE

Clustering algorithms based on ensemble called unsupervised ensemble learning or
consensus clustering can be considered as a modern clustering algorithm. Clustering
results are prone to being diverse across the algorithm, and each algorithm might work
better for a particular dataset. This diversity is hypothetically illustrated by a toy example
in Figure 4. In this figure, samples are in the same group represented by the same symbol.
As shown in figure, different clustering methods might give us different partitions of the
data, and they can even produce the different number of clusters because of the diverse
objectives and methodological foundations (Haghtalab, Xanthopoulos, & Madani, 2015).
As it will be discussed later, to deal with the potential variation of clustering
methods, one can use consensus clustering. The core idea of consensus clustering is to
combine good characteristics of different partitions to create a better clustering model. As
the simple logic of process is shown in Figure 5, different partitions ( , , … , ) need
2
1

∗
to be somehow produced and combined to create optimum partition ( ).