"If a conversation is about my research, I almost always have to explain what a graph is first. Unfortunately, it doesn't really help that the term is overloaded with so many meanings," Bojchevski says, laughing. The graphs he deals with are neither charts nor do they represent a mathematical function. When he speaks of graphs, he is referring to mathematical abstractions that can represent network-like structures such as human networks of relationships or protein-protein interactions.
If we imagine a social network such as Facebook, for example, its users would be represented in a graph as so-called nodes, which in turn are connected by edges. These edges express a relationship, for example, "is friends with." And so back to the small-world effects. These and other observed structural similarities and patterns in network structures help researchers today to build graph-based machine learning (ML) models. Some of these are already in use. Google, for example, uses such models in its Maps app to correctly determine the time of arrival at a destination, even in heavy traffic. But it's not just industry that benefits from artificial intelligence; graph-based ML models can also do amazing things in research. In 2020, for example, U.S. researchers used artificial intelligence to discover new antibiotics, including halicin, which has proven to be highly effective and can even combat bacteria, previously considered untreatable.
"If we want to use machine learning models in safety-critical areas and high-stakes applications such as medicine, then they must be absolutely trustworthy," says Bojchevski. This means that the models and the algorithms used for data processing must be robust. To explain what robustness means, the researcher again uses the example of social networks. Machine learning models in social networks can predict, for example, which user profiles are genuine and which are fake, says Bojchevski. The graph-based models look not only at the characteristics of the users themselves, such as age and gender but also at their relationships with other users - and those, too, can be meaningful. "The graph is never completely correct, however, because not all users reveal the same information about themselves, for example, or because attackers want to manipulate the graph to hide their own fake accounts. This is called noise in the graph, which can influence the prediction. If the graph is manipulated in a few places and the models collapse or the algorithms make false predictions, then they are not robust enough.To change this, on the one hand, Bojchevski takes on the role of an attacker and breaks the algorithms to expose their vulnerabilities. On the other hand, he assumes the role of a defender and tries to harden the algorithms to make them more robust. His latest research, however, bypasses the eternal cat-and-mouse game between attackers and defenders by developing models that are guaranteed to be robust.
Robustness is just one aspect of the trustworthiness of such models. "Trustworthiness also includes fairness, privacy, explainability, and uncertainty-awareness. I want to do a lot more research in these areas in the future," Bojchevski says, adding, "There are still a lot of problems in all of these areas today, partly because the decisions and predictions of complex models often can't be reproduced." The 30-year-old is convinced that the application fields of these complex ML models will nevertheless grow steadily and offer enormous potential.
Aleksandar Bojchevski, a native of northern Macedonia, completed his PhD on machine learning for graphs at the Technical University of Munich, where he had also previously earned a master's degree in computer science. He has been at the center since September and feels very comfortable here. "My research interests fit very well with the direction in which CISPA is currently growing. I immediately noticed in a positive way how much passion the researchers at CISPA have for what they are doing. CISPA has the potential to become something really big. I feel like I can still really help shape things here."
translated by Oliver Schedler
