Partenaires

CNRS
IN2P3
Sorbonne Universite
Universite de Paris
Initiative Physique des Infinis
UPMC


Rechercher

Sur ce site


Accueil > Thèses, Stages, Formation et Enseignement > Propositions de thèses antérieures > Propositions de thèses 2022 > Real-time rule induction in fraud detection and HEP

Real-time rule induction in fraud detection and HEP

by Tristan Beau - 26 April 2022

Title : Real-time rule induction in fraud detection and HEP

Supervisor : Mélissa Ridel

Co-supervisor : Bogan Malaescu

LPNHE Team : ATLAS

Description :

The goal of this PhD project is to automate the learning of a decision model by new combinations of statistical and knowledge based models applied to fraud detection and in high energy physics. Real time decision making combines today’s analytics and knowledge based models for fraud detection, notably in banking. Payment platforms detect in real time fraudulent transactions by combining recognition of human created patterns articulated on their banking symbolic knowledge model, and predictive models run to discover emerging fraud patterns by detecting new trends and anomalies from the data. In this thesis the student will work on new combinations of statistical and knowledge based models for a better decision automation in fraud detection and in high energy physics, for the recognition of human-created (fraud) and non-human-created patterns (in particle collision recorded with the ATLAS detector). While machine learning has been highly popular during the last years, their black box approach raises interpretability and explainability challenges. On the other hand, symbolic models, including rules, have been successful in making decisions more interpretable. Nevertheless they require to capture an existing knowledge or theory. In the context of real time decision automation, the student involved in this project will test the proposed numeric to symbolic model inferences to detect anomalies, patterns and anti-patterns, combining the efficiency of the numerical machine learning and the explainability of the symbolic approach. They will inject theory (knowledge from the Standard Model in physics, fraud detection patterns in financial transactions) and combine it with predictive models to classify observations and add an interpretability layer. Different angles in how we intend to combine numerical and knowledge models will be explored.

Web link :

[https://www.smarthep.org/positions/esr2/]

Workplace :

LPNHE, Paris and belgium

Facebook
Save this article in PDF