An empirical evaluation of random transformations applied to ensemble clustering

Abstract

Ensemble clustering techniques have improved in recent years, offering better average performance between domains and data sets. Benefits range from finding novelty clustering which are unattainable by any single clustering algorithm to providing clustering stability, such that the quality is little affected by noise, outliers or sampling variations. The main clustering ensemble strategies are: to combine results of different clustering algorithms; to produce different results by resampling the data, such as in bagging and boosting techniques; and to execute a given algorithm multiple times with different parameters or initialization. Often ensemble techniques are developed for supervised settings and later adapted to the unsupervised setting. Recently, Blaser and Fryzlewicz proposed an ensemble technique to classification based on resampling and transforming input data. Specifically, they employed random rotations to improve significantly Random Forests performance. In this work, we have empirically studied the effects of random transformations based in rotation matrices, Mahalanobis distance and density proximity to improve ensemble clustering. Our experiments considered 12 data sets and 25 variations of random transformations, given a total of 5580 data sets applied to 8 algorithms and evaluated by 4 clustering measures. Statistical tests identified 17 random transformations that are viable to be applied to ensembles and standard clustering algorithms, which had positive effects on cluster quality. In our results, the best performing transforms were Mahalanobis-based transformations.