Osman
Ali
Mian

Most machine learning researchers build systems that are very good at saying "what." Osman Ali Mian is interested in the harder question: "why." That distinction might sound academic. It isn't.

As a postdoctoral researcher at the Lamarr Institute and the Institute for Artificial Intelligence in Medicine (IKIM) at Ruhr University Bochum, Mian works at the frontier of causal discovery - the discipline of teaching algorithms to infer the actual cause-and-effect structure hiding inside data. Not correlations. Not patterns. Causes.

He does this with a practical obsession that is embedded in the title of his PhD thesis: Practically Applicable Causal Discovery. The thesis earned him a Magna Cum Laude distinction from Saarland University in February 2025. That's the German academic system's way of saying: this is exceptional work.

In January 2026, AAAI - the Association for the Advancement of Artificial Intelligence, one of the field's flagship conferences - agreed. Mian co-authored one of just five papers selected for the AAAI 2026 Outstanding Paper Award. The paper, "Causal Structure Learning for Dynamical Systems with Theoretical Score Analysis," introduced a method called CaDyT, which solves a stubborn problem: most causal discovery algorithms discretize continuous-time data, and that discretization introduces errors. CaDyT skips that step entirely, using a Difference-based Structural Causal Model that respects continuous-time dynamics from the ground up.

That framing - "predict responsibly" - is the key to understanding what drives Mian's work. Pure prediction is cheap. Models that memorize patterns can score well on benchmarks while failing catastrophically when the world shifts. Causal models don't just predict. They understand structure. They can say: if you change this, that will follow. That is a qualitatively different and more powerful kind of intelligence.

At AAAI 2026, Mian didn't just appear on the award podium. He had two separate papers accepted for oral presentation at the conference - a competitive distinction that puts his work among the top few percent of submissions. The second oral, "SEQRET: Mining Rule Sets from Event Sequences," tackles pattern discovery in sequential data - different tools, same underlying commitment to interpretable, structured reasoning.

Here is the problem with most AI today: it can spot patterns across millions of examples, but it doesn't know which patterns matter under intervention. A model trained to diagnose disease in hospital A might fail in hospital B - not because the disease changed, but because the data-generating process did.

Causal models are built differently. Instead of learning "X predicts Y," they learn "X causes Y." That's harder. It requires assumptions, mathematical identifiability conditions, and careful algorithmic design. But the payoff is a model that keeps working when the world changes, because it understood the mechanism, not just the correlation.

Mian's contributions span the full stack of this challenge. At the observation layer, his AAAI 2021 paper, "Discovering Fully Oriented Causal Networks," tackled the fundamental problem of orienting causal edges from data. At the inference layer, his ICML 2022 work addressed cause-and-effect identification under heteroscedastic noise - the kind of messy, variable-variance noise that real data generates. And at the systems layer, his CaDyT work handles dynamical systems where the causal relationships evolve through time.

Each paper is a tool. Together, they form a toolkit for applying causal reasoning where it matters: in healthcare, in federated industrial systems, and in any setting where getting the causal story wrong has real consequences.

His AISTATS 2026 paper pushes further still, proposing a unified framework that handles causal discovery and data imputation simultaneously - two problems usually solved in sequence, now solved together. The practical target is medical records, where missingness is not random and imputation errors compound.

The title of Mian's PhD thesis wasn't an accident. "Practically Applicable Causal Discovery" is a research agenda, not just a dissertation chapter. His promotion committee - four professors across three institutions, including causal inference heavyweights Murat Kocaoglu and Jilles Vreeken - validated that the work delivers on that promise.

Now, as a postdoctoral researcher, the stakes are higher. IKIM - the Institute for Artificial Intelligence in Medicine - is not an abstract research environment. It sits inside a university hospital ecosystem where the algorithms eventually touch real patients. Federated learning, which lets institutions collaborate without sharing raw data, is not an academic convenience in that context. It's a legal and ethical requirement.

Mian's trajectory from CISPA to Amazon to IKIM is a straight line: build methods that work under the constraints of the real world, not the idealized constraints of benchmark datasets. Privacy. Heterogeneity. Dynamical systems. Missing data. Each constraint is a paper. Each paper is a step toward AI that can be trusted with decisions that matter.

The Outstanding Paper Award at AAAI 2026 is a signal. The field is watching.