How Reliable Are Machine Learning Methods? With Anne-Laure Boulesteix and Milena Wünsch

Often a new machine learning method claims to outperform the last. Whether it’s in bioinformatics, finance, or image recognition, the message is the same: this algorithm is faster, more accurate, more powerful. But can we trust those claims?

Beneath the surface of many benchmarking studies lies a quiet problem: subtle biases that skew comparisons and inflate performance. These issues often go unnoticed — but they can have real consequences, especially when such models are used to inform research or high-stakes decisions.

Anne-Laure Boulesteix, Professor of Biometry at LMU and MCML PI, and Milena Wünsch, PhD student at LMU and MCML, study how seemingly harmless methodological choices can lead to misleading results.

One common issue: when a method fails on a dataset, researchers may simply drop it from the analysis. While convenient, this can introduce bias and overstate performance.

Bias can also arise from less obvious sources — like spending more time tuning one method, being more familiar with a tool, or unconsciously interpreting results in its favor.

With so many studies promoting the “next best” algorithm, it’s hard to know which results to trust. Researchers may end up using a method that only looked good due to biased comparisons. Still, the researchers are hopeful. In recent years, the methodological machine learning community has made real progress — pushing for better standards, more transparency, and more careful benchmarking.

The film was produced and edited by Nicole Huminski and Nikolai Huber.