Machine learning to deal with uncertainty in knowledge base for multivariate clustering applied to spatial analysis

We present a Tailor Made Machine Learning (TMML) methodology combining different clustering algorithms, spatial statistical methods and cartographic tools. The methodology is currently being programmed in an R package, especially designed to handle multivariate spatial datasets. We highlight the strengths of unsupervised clustering for the management of environmental health phenomena, also pointing out several uncertainty sources affecting the results of our analysis. In particular, we acknowledge that the traditional hierarchical clustering is usually applied without performing dynamic reallocations, integrating spatial key-concepts or discussing the quality of outputs. Therefore we describe the foundations of the TMML methodology, which is applied to deal with these uncertainties, as well as with the variety of possible outputs. The R functions are applied to the spatial dataset included in the package so to illustrate the procedure to apply for identifying the most accurate clustering output, in the context of a sustainable agriculture example in Luxembourg.