Machine Learning Algorithms and Explanations to Detect Banking Fraud
This internship is a collaborative project between BPCE group (Banque Populaire Caisse d'Epargne) and the artificial intelligence institute of Toulouse (ANITI).
It will be followed by a CIFRE thesis with the same partners .
The internship will be based in BPCE SA (member of the Communauté BPCE) in the Secrétariat Général department, which deals with compliance, legal, permanent controls, security, public relations, etc.
The team that will host the intern is called Intelligence Artificielle & Modèles and has for mission to develop AI systems (Machine Learning algorithms, data science studies, robotization of simple processes) for the benefit of the rest of the Secrétariat Général.
The academic partner is the Artificial and Natural Intelligence Toulouse Institute (ANITI) and in particular the group DeepLEVER (for Deep Learners Explanation and VERification), which aims at explaining and verifying machine learning systems via combinatorial optimization in general and SAT in particular.
Joao Marques-Silva and Emmanuel Hebrard will be the academic co-advisors. The intern will also have access to LAAS-CNRS to have regular meetings the academic supervisors.
Industrial Problem :
BPCE is interested in designing Machine Learning models, in particular to detect illegal activities such as fraud (involving checks, money transfers or card payments), dirty money laundering or terrorism financing.
Moreover, the models should satisfy a range of properties :
In particular, this rules out deep learning or other complex systems, and the focus will be on rule-based models (such as decision trees, decision lists, decision sets, etc.)
Therefore, the models should be rather stable but adaptive to new behaviors.
Scientific Objectives and Research Project :
The goal of this thesis is to design efficient solutions meeting all the criteria described above.
In order to be interpretable, rule-based or logic-based machine learning models will be the main focus. A large part of the research work will be devoted to analyse and compare the different models (decision trees, decisions sets, Boolean decision diagrams, etc.) in light of :
Then, the research will focus on the design of efficient methods to learn, explain and adapt the chosen model(s). The approaches can be either algorithmic or based on optimisation paradigms (Boolean Satisfiability, Constraint Programming, Integer Linear Programming, etc.
Finally (and possibly) new machine learning models better adapted than existing ones may be proposed.