Clustering ensembles

In Chapter 3, Predicting Sports Winners with Decision Trees, we looked at a classification ensemble using the random forest algorithm, which is an ensemble of many low-quality, tree-based classifiers. Ensembling can also be performed using clustering algorithms. One of the key reasons for doing this is to smooth the results from many runs of an algorithm. As we saw before, the results from running k-means are varied, depending on the selection of the initial centroids. Variation can be reduced by running the algorithm many times and then combining the results.

Note

Ensembling also reduces the effects of choosing parameters on the final result. Most clustering algorithms are quite sensitive to the parameter values chosen for the algorithm. Choosing slightly different parameters results in different clusters.

Evidence accumulation

As a basic ensemble, we can first cluster the data many times and record the labels from each run. We then record how many times each pair of samples...