Learning Deep Representations for Class Frequency Estimation

• Applicant: Dr. Mirko Bunse
• Subject Area: Methods in Artificial Intelligence and Machine Learning
• Term: since 2026
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 576571177

This proposal addresses a key challenge in supervised machine learning: estimating varying class frequencies in multiple large datasets rather than the usual classification of individual data points. Solving this challenge is critical in diverse application areas like physics, the social sciences, public health, and environmental monitoring, where understanding the global frequency of certain classes (e.g., particle types, opinions, insect or plankton populations) is more important than classifying individual instances. Class Frequency Estimation (CFE) has become a specialised area of machine learning research, including tailored methods and evaluation measures. Prior work by the applicant highlights that existing CFE algorithms mainly differ in their data representations, while other aspects, like loss functions, often depend on application-specific requirements. Other researchers have made further progress in characterising optimal data representations under the specific condition of prior probability shift. However, two major challenges remain: (1) how to learn optimal data representations from finite training sets and (2) how to handle complex dataset shifts beyond the standard assumption of prior probability shift. These unresolved issues limit the broader applicability of current CFE methods despite a pending demand for accurate class frequency estimates. This project aims to address these gaps by developing robust CFE methods through optimised CFE data representations. The approach will involve novel deep model architectures and representation-specific loss functions designed to handle severe and complex dataset shifts. Success will be ensured through empirical benchmarks and fundamental theories for the rigorous evaluation of method performance, promising significant benefits across the diverse application areas of CFE.

• DFG Programme: Research Grants
• International Connection: Brazil, Italy, Spain
• Co-Investigators: Professor Dr. Jakob Rehof; Professor Dr. Wolfgang Rhode
• Cooperation Partners: Dr. Pablo González; Dr. André Maletzke; Dr. Alejandro Moreo Fernandez; Dr. Fabrizio Sebastiani