Big Data Analytics and Cognitive Computing

Recent advances in computer, communication and sensor actuator technologies hint at emerging information paradigms, among which Big Data appears as the most prominent. Big Data offers the capability to collect (notion of “data marketplaces”), combine, cross-analyze, and process large volume of monitoring and operational data, using high performance and cloud computing infrastructures. In terms of application, Big Data promises are potentially extremely large.

In that context, IRIXYS members initiated in 2015 in partnership with the Worldline company program aiming at investigating the potential of prescriptive analytics for fraud detection. Prescriptive analytics, which is a relatively new field in the Big Data analytics, goes beyond descriptive and predictive analytics. It uses a combination of techniques and tools such as business rules, optimisation and simulation algorithms, machine learning and computational modelling procedures to advice on possible outcomes. The main originality of the program with Worldline originality is to use Linked Data and Linked Open Data to semantically enrich processed sets of data, and improve decisions processes. This program constitutes one of the scientific pillars of the Centre for the next 5 years.

In the next years, IRIXYS will pursue its activities in the field of Big Data Analytics incorporating progressively Cognitive Computing.  Cognitive Computing handles human kind of problems. It aims to develop in a computerized model, a coherent, unified, universal mechanism simulating human thought processes. Cognitive computing is used in numerous artificial intelligence applications, including natural language programming, neural networks, robotics and virtual reality.

Research works on this topic are listed below:

  • Fraud Detection using Machine Learning with Integration of Contextual Knowledge

We aim to detect fraudulent transactions in a real-world database. The context of this work is credit card fraud detection. Credit card transactions depend on a temporal context (e.g., calendar events). Also, there are dependencies in the sequence of transactions of a card holder. We want to integrate this contextual knowledge in a classifier used to detect frauds on a real world dataset.

Franco-German PhD thesis of Yvan Lucas, supervised by Prof. Sylvie Calabretto and Prof. Michael Granitzer, co-supervised by Dr. Léa Laporte, Dr. Pierre-Edouard Portier in collaboration with Atos Worldline.

  • Multi-Criteria Decision Making in Security, based on a large number of heterogeneous sources

Using the target Credit Card Fraud detection as a case study benchmark, we aim to investigate on theoretical tools and methods to improve over the current state of the art in term of: aggregated classifier performance (recall, precision, …) and performance oriented rule management (keeping in the core pool, for the stream processing, only the most important / effective; presenting to the online investigators the K top-suspect transactions, based on activated rules importance).

Franco-Italian PhD thesis of Léopold Ghemmogne Fossi, supervised by Dr. Gabriele Gianini and Prof. Lionel Brunie, co-supervised by Prof. Ernesto Damiani and Dr. Sonia Ben Mokhtar in  collaboration with Atos Worldline.

  • Learning Feature representations for Large Graphs
Mining the information available in the large social networks usually needs to learn useful network representations. An effective way is to embed the network into a low-dimensional space which preserves the structure of the original network. The goal of our research work is to learn vector representations of nodes in a large graph using neural networks.

Keywords: Graph Embedding, Neural networks, Representation Learning

PhD thesis of Fatemeh Salehi Rizi, supervised by Prof. Michael Granitzer.