Unsupervised machine learning approach for building composite indicators with fuzzy metrics

This study aims at developing a new methodological approach for building composite indicators, focusing on the weight schemes through an unsupervised machine learning technique. The composite indicator proposed is based on fuzzy metrics to capture multidimensional concepts that do not have boundaries, such as competitiveness, development, corruption or vulnerability. This methodology is designed for formative measurement models using a set of indicators measured on different scales (quantitative, ordinal and binary) and it is partially compensatory. Under a benchmarking approach, the single indicators are synthesized. The optimization method applied manages to remove the overlapping information provided for the single indicators, so that the composite indicator provides a more realistic and faithful approximation to the concept which would be studied. It has been quantitatively and qualitatively validated with a set of randomized databases covering extreme and usual cases. [Display omitted] •Fuzzy metrics allow benchmarking and the assessment of progress towards set targets.•The composite indicator is computed by using unsupervised machine learning techniques.•For calculating the importance of single indicators a optimization model is proposed.•The robustness is checked conducting simulations.•We fill the gap in traditional methods of building composite indicators.