Feature Selection for Cluster Analysis in Spectroscopy

Cluster analysis in spectroscopy presents some unique challenges due to the specific data characteristics in spectroscopy, namely, high dimensionality and small sample size. In order to improve cluster analysis outcomes, feature selection can be used to remove redundant or irrelevant features and reduce the dimensionality. However, for cluster analysis, this must be done in an unsupervised manner without the benefit of data labels. This paper presents a novel feature selection approach for cluster analysis, utilizing clusterability metrics to remove features that least contribute to a dataset's tendency to cluster. Two versions are presented and evaluated: The Hopkins clusterability filter which utilizes the Hopkins test for spatial randomness and the Dip clusterability filter which utilizes the Dip test for unimodality. These new techniques, along with a range of existing filter and wrapper feature selection techniques were evaluated on eleven real-world spectroscopy datasets using internal and external clustering indices. Our newly proposed Hopkins clusterability filter performed the best of the six filter techniques evaluated. However, it was observed that results varied greatly for different techniques depending on the specifics of the dataset and the number of features selected, with significant instability observed for most techniques at low numbers of features. It was identified that the genetic algorithm wrapper technique avoided this instability, performed consistently across all datasets and resulted in better results on average than utilizing the all the features in the spectra.