Mert Kosan

Professional Focus

As a Staff Machine Learning Scientist at Visa, I apply these interests to fraud detection, risk modeling, and payment intelligence at global scale. My work includes the design of large-scale self-supervised transformers and Large Transactional Models for multi-task fraud and risk applications, as well as a GenAI-powered explainability system that generates human-readable rationale for high-risk transaction decisions — supporting the kind of transparent, human-in-the-loop decision-making that I believe should underpin any consequential AI system. I have also designed the feature and signal architecture for an account-level insights system operating across billions of accounts, and I regularly mentor junior scientists and engineers on both technical direction and research methodology.

I place equal value on methodological rigor and real-world deployment. A well-designed method that never reaches production, and a deployed system that cannot be understood or trusted by the people relying on it, represent, to me, two versions of the same failure. Much of my work is concerned with closing that gap.

Research Interests

My research centers on transparent representation learning, with a particular focus on graph-structured data and human-AI collaboration. During my Ph.D. at UC Santa Barbara, advised by Prof. Ambuj K. Singh, I worked on formulating explainability not as an auxiliary feature of machine learning systems, but as a core design constraint. This work produced GCFExplainer, a method for generating global counterfactual explanations for graph classification, and GNNX-Bench, a comprehensive benchmarking study of graph neural network explainers — recognized among the Top 10 papers at WSDM 2023 and awarded Best Paper at MLoG-WSDM 2023.

I am drawn to problems that are not yet well-formalized — questions at the boundary of existing methodology, where the first task is determining how to think about the problem before attempting to solve it. I find this kind of open-ended inquiry, followed by rigorous formalization, to be the most intellectually rewarding part of research.

Background

I hold a Ph.D. and M.S. in Computer Science from the University of California, Santa Barbara, and a B.S. in Computer Science from Sabancı University, where I graduated first in my program. My research has been published at venues including KDD, ICLR, AAAI, and WSDM, and I hold three patents in machine learning applied to fraud detection and anomaly analysis. I currently serve as an Area Chair for KDD's Dataset and Benchmark Track and have reviewed for venues including NeurIPS, ICLR, and The Web Conference.