Unsupervised learning, especially graph-based Manifold Learning (ML) dimensionality reduction methods have great importance in today’s data analysis studies. These methods are basically defined as spectral projection of the high- dimensional data to a lower dimensional space while preserving its geometric properties. However, these methods might cause some realization problems in big data analysis due to their high computational complexity. The reason of this is because the ML methods are often used as a black-box function with their embedded linear algebra building blocks, that are not convenient to high performance computing platform. On the other hand, modern applied linear algebra studies provide several different methodologies to solve large-scale systems efficiently. Another critical open-problem in the ML methods is the determination of intrinsic dimensionality. In this project, a novel scalable high-performance computing method based on spectrum slicing techniques will be developed as applied to Laplacian Eigenmaps for solving generalized eigenvalue problem. This new technique will also have an ability to automatically determine the intrinsic dimensionality without consuming any extra computational resource.
E. Fatih Yetkin (Kadir Has University)
Links to the reports
Links to the datasets used in the study
We are currently seeking talented MSc and PhD students to join our project. If you are interested, please send your CV to E. Fatih Yetkin
“ ”