Indefinite proximity learning for multi-view structured data
Machine learning increasingly shapes our daily lives, from social media recommendations to environmental sciences or medical diagnoses. However, many real-world datasets—such as DNA sequences, social network connections, or patient records in a hospital—have complex structures that modern popular techniques like deep learning cannot effectively handle when data is limited. In his thesis, Maximilian Munch improves classical and creates new methodologies to analyze such structured data more effectively.
Commonly, traditional approaches force data into simplified mathematical forms, often losing important information in the process. Instead, Munch embraces the natural complexity of structured data, developing smart transformation techniques that preserve essential patterns while making the data analyzable by computers. In general, Munch introduced two main innovations. First, Munch describes improved methods for processing individual complex datasets, making analysis faster and more accurate while using less computer memory. These techniques work particularly well when data is scarce—a common challenge in medical research and other specialised fields. Second, Munch developed new approaches for combining multiple perspectives of the same data.
Similar to how viewing an object from different angles provides a better understanding in real life, analysing data from multiple viewpoints yields richer insights. The developed fusion methods intelligently combine these perspectives, achieving better classification accuracy than existing techniques. By enabling more effective learning from limited, structured data, this work expands traditional machine learning capabilities, making sophisticated data analysis accessible for real-world scientific and industrial challenges.