**5. Results**

## *5.1. Feature Selection*

The computational complexity of the Wasserstein distance can quickly become intractable in the case of multi-variate histograms, as already mentioned. The computation of the barycenter and performing the clustering procedure using the WST distance add substantially to the computational cost. It is therefore important to reduce the number of variables to consider, and for this reason, a feature selection strategy based on the Information Gain (IG) is used to select the most relevant KPIs. In turn, each KPI is considered as a target variable in a classification problem, and the IGs of all the others KPIs are computed. Since seven KPIs are considered, for each of them, six different values of IG are obtained, each of which represents the importance for the specific KPI in predicting the other six. Therefore, for each KPI, the average of these six represents its IG. Table 3 reports these results. In the following analysis, the four most relevant KPIs are considered.

**Table 3.** Information Gain of the seven KPIs.

