Master thesis : Exploring Antitrust Cases with Clustering Methods

Abstract

The use of artificial intelligence in today's world has been growing for many years in many fields such as law. However, a part of law, called competition law or antitrust law, has been put aside. Hence, no machine learning or intelligent system helps antitrust law enforcers in their day-to-day work. This thesis seeks to fill this gap in order to determine whether it is possible not to automate the decision-making aspect of the antitrust judges' job. The goal is to know to what extent automation is possible thanks to AI applications in the antitrust field. Can a sentencing decision be taken by artificial intelligence systems? If not, is it possible to guide the judge by providing him patterns identified from older legal cases?

The first part of this thesis presents the various intersections between law and artificial intelligence. Then, there will be an explanation of what antitrust law is. This first part will end with the related works of the clustering methods used.

The second part is the one including the more technical aspects of the thesis. Firstly, the database and the different modifications made on it beforehand will be developed. Then, the different methods used to compute the performances of the algorithms will be presented. The different clustering algorithms will also be explained and analyzed. Moreover, several feature selection techniques will be developed and tested to determine the most relevant features. This part will conclude by determining that K-Means after the SPEC feature selection technique is the solution giving the best performances.

The last part presents a more legal analysis of the clusters formed using the most efficient methods for the available database. Indeed, the initial question which is to automate as much as possible the decision process of a judge, must be answered. In this part, the similarities of the legal cases within the same cluster will be put forward in order to prove that patterns exist and that the clustering method has allowed to determine them.