Pairwise Constrained Clustering and Robust Regression: A Case Study on French Enterprise Activities and Expenses Data
Ines Akaichi, Patrice Wislez · 2021
Argedis, a subsidiary of Total France in charge of the management and operation of part of its network of service stations. Argedis seeks to better control its expenses and its annual budget by modeling the station loads using regression models. The company believes that the service stations have different profiles depending on the type of activity, which has an impact on the level of expenses. To obtain accurate predictive models, we used the results of grouping the service stations using a semi-supervised clustering algorithm. The data describing the activities of the company show irregular values. Therefore, we choose robust regression models to remedy the irregularity detected in the data. The results obtained show that the removal of outliers improves the metrics of the models. However, other types of outliers are deemed necessary to be preserved in the data modeling steps.