Detection of Anomalous contracts of the Italian Public Administration

The Italian laws enforce the public entities, such as municipalities or general government agencies and public authorities, to provide the Italian National Anti-Corruption Authority the information about each public procurement contract issued. The collection of public contracts enables the extraction of information relevant for companies in the market of public procurement. From the need of offering reliable information to such companies arises the need of data cleaning and anomaly detection. This thesis' project provides an outlier detection tool with the aid of machine learning techniques. Initially, the relations between the variables of the dataset are analyzed and discussed. Then a baseline model comprising of a set of heuristics is developed with the aid of a domain expert. The resulting outliers are the ground truth for comparing three other models: kernel density estimator, Gaussian Mixtures and one-class support vector machine. The results shows that the KDE with a normal scale bandwidth selector exceeds the baseline performance.