Characterization of the k-means algorithm for spectral profiles

Abstract

The k-means algorithm is a Machine Learning clustering method that has gained popularity both for its scalability and its simplicity. The output of this method contains a distribution of the input data in k groups as well as k representative examples. The aim of this Bachelor’s Thesis is to test k-means clustering results under controlled conditions by means of an artificial dataset. The data mimic solar observations from the Interface Region Imaging Spectrograph (IRIS) in the Mg II h&k lines. The situation is made incrementally more complex and the impact on the clustering is studied on a case by case basis. The goal is to consistently obtain a distribution that accurately separates the different profiles in the dataset. Furthermore, the results are compared to those of hierarchical clustering methods and the effect of two common preprocessing schemes is analyzed. The k-means final results are considered satisfactory, given that the main goal of discerning between spectral behavior patterns is achieved with very low error rates, even when the data are purposefully contaminated with defective profiles and noise. Nevertheless, when these impediments become too widespread, masking becomes necessary, allowing for the previous statistics to be recovered. The hierarchical methods are deemed equal or inferior to k-means in terms of performance, depending on the specific criterion.