Evaluación de algoritmos de aprendizaje de máquina no supervisados con datos climáticos

Abstract

When using climate data, researchers have difficulty determining the clustering algo-rithm and the best performing parameters for processing a specific dataset. We evaluated of the following unsupervi-sed machine learning algorithms: K-means, K-medoids and Linkage-complete, which are applied to three datasets with clima-tological variables (temperature, rainfall, relative humidity, and solar radiation) for three meteorological stations located in the department of Caldas, Colombia, at diffe-rent heights above sea level. Five scenarios are defined for 2, 3, and 5 clusters for each of the two partitioned algorithms, and five scenarios for the hierarchical algorithm, in each one of the meteorological stations. Different quantities and groupings of varia-bles are applied for the different scenarios by using Euclidean distance. Davis-Bouldin is the applied method of quality evaluation of clusters. Normalization with techniques such as range-transformation and Z-trans-formation, as well as some iterations of the algorithm and reduction of dimensionali-ty with PCA. In addition, the computatio-nal cost is evaluated. This study can guide researchers on certain decisions in cluster analysis used in meteorological data, as well as identify the most important algorithm and parameters to take into consideration for the best performance, according to par-ticular conditions and requirements.