Media outlet: CMU Portugal

Maria da Loura Casimiro is a CMU Portugal Dual Degree Ph.D. student in Software Engineering at INESC-ID/Instituto Superior Técnico and the Software and Societal Systems from Carnegie Mellon University School of Computer Science. She started her Ph.D. in 2019 shortly after completing the MSc. in Electrical and Computer Engineering. This year, she presented her current work on “Self-Adaptive Machine Learning” to Feedzai under a meeting promoted by CAMELOT – a CMU Portugal large project promoted by Feedzai, where Técnico is one of the partner institutions.

As a result of this meeting, she was invited by the company’s Chief Science Officer to do a 3-month summer internship to develop her ideas by leveraging a real industrial use case, an opportunity that she would not miss: “I had already discussed with my advisors the possibility of doing an internship at a company, not only because this was something that I was curious about and interested in doing but also because it would contribute to the fulfilment of one of my requirements at CMU. So, when I had the opportunity to do the internship at Feedzai, I was delighted and very excited. Feedzai is a rising company in the fraud detection domain (one of Portugal’s unicorn companies), I was going to be exposed to cutting-edge technology, and the topic of the internship was very much aligned with my PhD topic.”

Feedzai is one of the first startups founded under the scope of the CMU Portugal Program back in 2010. Nowadays, the company is the market leader in fighting financial crime, providing today’s most advanced cloud-based risk management platform powered by artificial intelligence and big data, which involves a constant need to improve its models and keep up with evolving financial frauds. According to Maria Casimiro, “Feedzai’s core business is not explicitly related to the work I’m developing. However, the technologies they leverage and the context in which they are used, constitute a possible use case for the work I’m developing, which is related to improving the operation of Machine Learning (ML) based systems, i.e., systems that rely on one (or more) ML components to operate”.

The project we assigned to Maria was a great challenge both for her and for Feedzai. It was completely uncharted territory, a new research line (Marco Sampaio)

At Feedzai, she was supervised by João Ascensão, Senior Director of Research and Marco Sampaio, Advanced Data Scientist. Marco referred to this experience as a challenge, adding that “the project we assigned to Maria was a great challenge both for her and for Feedzai. It was completely uncharted territory, a new research line.” Maria was involved at all stages. From defining the goals of the project, reviewing state of the art, and defining the detailed experimental setup. Maria kept a positive attitude from day 1, despite the great challenges she had to face since this was a very new experience. Namely, dealing with extensive industry datasets, a new use case (fraud detection), and very different and specific evaluation goals.”

Regarding the relevance of her research work to the company, Marco Sampaio explains that “Maria has been focusing on problems related to optimizing machine learning jobs usually related to parameter searches in a high dimensional space. This is a task that is relevant for most companies that leverage machine learning as part of their products, which is Feedzai’s case. The fact that Maria was familiar with these topics was relevant to start working on the project.”

On her side, Maria stated that the main goal of developing her ideas by leveraging a real industrial use case was highly achieved. She reinforces this idea that “By learning more in detail about how fraud detection systems operate and how fraud detection is performed, I was able to formulate better the research problems I was looking into and have specific examples from the fraud detection domain. Having these examples is extremely relevant to motivate the need for the work I’m developing and to communicate the challenges/needs to other researchers.”

This allowed me to understand how research is done outside of Academia. It was very fascinating to see how the whole team works together (Maria Casimiro)

Regarding this first professional experience in a company, she referred to it as a very positive experience that she highly recommends to other colleagues “I was fortunate to have the chance to do my internship with the research team at Feedzai. This allowed me to understand how research is done outside of Academia. It was very fascinating to see how the whole team works together and to understand the process of selecting the research projects that are expected to be the most successful. I learned a lot about what my strengths and weaknesses are. I learned new concepts and how to work with new platforms. For these reasons, I would advise my colleagues to have similar experiences.”

These kinds of internships are an important part of the company’s R&I strategy and are “often useful to try high-risk ideas in a time scoped manner, through simple experiments. This is very useful to diversify our projects and provide information on interesting new avenues to pursue”, explains Marco Sampaio.

This summer internship resulted from Maria’s connections established during the first two years of her Dual Degree Ph.D. Now into her third year, she describes the experience as challenging but very rewarding “I’m learning a lot about different areas of research, and specifically about the one I’m working in, but I’m also learning a very much about myself, and about what I want after I graduate.”

Maria Casimiro spent her second year at CMU, an experience that considered to be “eye-opening”, “From an academic perspective, it was very interesting to see how the requirements for Ph.D. students and what is valued by both professors and students differ between IST and CMU; how the interactions between Ph.D. students and professors are much more informal; and how there are always so many initiatives happening (e.g. talks). Additionally, I really enjoyed the courses I took there. The professors were very passionate about what they were teaching, and the classes were extremely engaging. From a personal perspective, it was very interesting to meet so many people from different backgrounds and to learn about American culture. One experience I will never forget is the Thanksgiving holiday, which I spent with a friend from CMU and her family. Being at CMU was an awesome experience and I’m glad I still have the chance to go back again.”

And plans for after graduation? “After graduation, I see myself embarking in Industry, perhaps doing research. Right after graduating, I would like to work abroad for some time, to gain more expertise, but I hope to stay in Portugal in the long run.”

AIDA is a CMU Collaborative Large Scale project launched in 2020 with the main goal to improve the RAID platform used by Mobileum for integrated risk management in companies. To achieve this objective, the Portuguese ICT Company that is currently the leading provider of Telecom analytics for roaming, security and risk management joined INESC TEC, Universidade de Coimbra and the Computer Science Department at Carnegie Mellon University (CMU) to apply to the  CMU Portugal Call for Large Scale Collaborative Research Projects (LSCRPs).

After a thorough evaluation process, AIDA was one of the 12 research projects selected for 3 years’ funding, but the project will expectably end before that.

“The evolution of the Telecom world made us rethink the approach in terms of the analytics strategy.” (Carlos Martins)

According to Carlos Martins, Head of Engineering at Mobileum, “the evolution of the Telecom world made us rethink the approach in terms of the analytics strategy. With the evolution of the Telecom business, there was a significant shift from a process to a functional orientation. Business is now fully data-driven in a streamed approach with a decoupled analysis from any data aggregation and demanding schemas-on-read instead of pre-defined schemas.”

The RAID platform handles enterprises’ entire risk management lifecycle and is currently applied worldwide to address, among others, revenue assurance, business assurance, and fraud management. The AIDA project’s overarching goal is to conceive a new version of the RAID platform by moving some of the pipeline phases to the edge of the system and taking advantage of 5G. Today, the platform is fully deployed in physically collocated servers, either on-premises or in the cloud. With AIDA, data collection, monitoring, and even actuation phases should be prepared to run in diverse hardware architectures outside the platform’s owner’s physical or even administrative control.

According to Carlos Martins “5G will promote a massive growth of data. Part of the 5G protocol also permits multi-access edge computing (MEC), enabling wireless operators to offer a high level of automation from the distributed ML and AI architecture at the network edge. Fraud detection is nowadays mostly addressed by ML and AI, and AIDA will help RAID be ready to do the fraud detection at the edge level, ensuring quicker detection time and minimizing data transfer to a central cloud, while ensuring full data encryption on any of the needed data transfer.”

Thanks to this newest version of the platform, companies will be able to collect and monitor data in a highly flexible way, with real-time guarantees, security and reliability by leveraging the increasing edge computing capacity made available by the IoT and the imminent large-scale deployment of 5G cellular technology.

The goal of the project is clear as is the expected impact on Mobileum business: “From a qualitative point of view, we expect to bring a higher ROI for our customers by being able to address new fraud detections and corresponding preventive actions. From a quantitative point of view, we expect to acquire new customers that are now doing the path to 5G, and that this project will enable us to address it with strong competitive advantages” says Carlos Martins.

Christos Faloutsos, Professor at CMU’s Computer Science Department, leads the research team in Pittsburgh and supervised the work done during this first year. He adds that “at CMU we managed to get access to (anonymized) phone records. From here we have already started developing algorithms to find patterns, anomalies, and fraudsters. My group is working on the algorithms and visualization aspects of mining such huge amounts of complex data (who-calls-whom-and-when). There are fascinating challenges, both from the machine learning point of view (like “do textbook assumptions of Gaussianity independence hold?”), from the visualization point of view (“how to plot a graph with millions of nodes”), and from the systems point of view (how to store, and access, these large datasets).”

Regarding the benefits of the collaboration established under AIDA, which involves Mobileum as promoter and researchers both in Portugal and CMU, Faloutsos is confident in stating that “there are immense benefits, for both sides. My group and I are discovering unexpected patterns in the real graphs, which helps us design better algorithms. My students get hands-on experience with real data, with all the surprises that textbooks don’t mention (missing values, occasionally wrong labels, deviation from typical distributions like the Gaussian one, and many more). The domain-expertise that Mobileum provides, is extremely valuable: it helps us understand what are the most important business problems that we, scientists, should focus on, and helps us translate them into mathematical problems that we can subsequently solve. Conversely, Mobileum is benefiting from state-of-the-art algorithms in the literature (several of which are from CMU), and they will be the first to benefit from the new algorithms that my group is developing.”

On the Portuguese side, Carlos Martins also has high expectations for this research partnership “we expect a lot from these world-class researchers. We followed the work of some of the CMU colleagues in the last years and having now the chance to work daily with them is really beneficial to both – we learn from new techniques while challenging CMU with new use cases on Telecom Fraud. At the end of the day, I believe that both parties are learning a lot from each other.”

“The marriage between scientific research and production-grade technology is always the biggest challenge” (Carlos Martins)

Even if the project’s future seems bright, there are always some challenges to overcome. “The marriage between scientific research and production-grade technology is always the biggest challenge, so there are some technologies/techniques that still haven’t prove it’s feasibility in a real-world scenario. Additionally, data is the new oil, and having real-world data in a scientific approach is sometimes a challenge. Synthetic datasets are being explored, still, real-world data will be the key driver for project acceptance”, states Carlos Martins.

By the end of the project in 2022, the objective is to deploy a complete prototype of AIDA in a realistic scenario in telecommunication services. According to Carlos Martins, this prototype will be “a very resilient and flexible system that will detect Fraud in a Multi Edge Cloud architecture, dazzling the users with great accuracy and finding new Fraud patterns that were not possible before – we have great expectations on the results of new explored techniques like GNNs and Metric Forensics.”

After the end of the AIDA project, and based on the success of this partnership, Christos Faloutsos leaves the door widely open for future collaborations “the collaboration has been extremely productive and pleasant, with weekly remote meeting with Mobileum (Pedro Fidalgo and his group) and the main advantages as mentioned before. My group and I would be delighted to maintain these collaborations for as long as possible.”

More on the AIDA project at: https://cmuportugal.org/large-scale-collaborative-research-projects/aida/