Information
- Publication Type: Master Thesis
- Workgroup(s)/Project(s):
- Date: 2025
- TU Wien Library: AC17416388
- Open Access: yes
- First Supervisor: Manuela Waldner
- Pages: 62
- Keywords: Dimensionality Reduction, Data Visualization, Prototypical Network, ProtoNet, Class Separation, Scatter Plot
Abstract
Making sense of data is something that many professionals are required to do on a daily basis. This can be a difficult task if the amount of data is so large that it can not be easily examined. One effective method of quickly getting an overview of data structure is visualization, but this is not always a feasible solution with large data due to the sheer amount of data and also the potentially high dimensionality. Machine learning models can help with with the organization and classification of data, but they often require large quantities of labeled training data, which is frequently not readily available. This is why models that can reliably classify data based on only few examples for each class are an interesting topic of research. One such kind of model are prototypical networks. They utilize few samples to create an embedding space in fewer dimensions, in which similar data points cluster around a single class prototype. In this thesis, we investigate if the embedding space of a prototypical network makes for a good approach for the purpose of visualizing high-dimensional, unstructured data. The goal is to reduce the dimensionality of the data in such a way that the highdimensional relations and structures between data points are preserved, resulting in 2D representations of the data that form coherent class clusters in a scatter plot visualization. This approach is compared with, and evaluated against, other well known supervised and unsupervised dimensionality reduction techniques. Through quantitative experiments relying on statistical measures, as well as a qualitative evaluation of our results, we find that our ProtoNet is capable of producing point embeddings in which the spatial separation of classes is as good or better than the other methods.Additional Files and Images
Weblinks
BibTeX
@mastersthesis{stoff-2025-pvu, title = "Prototypical Visualization: Using Prototypical Networks for Visualizing Large Unstructured Data", author = "Mario Stoff", year = "2025", abstract = "Making sense of data is something that many professionals are required to do on a daily basis. This can be a difficult task if the amount of data is so large that it can not be easily examined. One effective method of quickly getting an overview of data structure is visualization, but this is not always a feasible solution with large data due to the sheer amount of data and also the potentially high dimensionality. Machine learning models can help with with the organization and classification of data, but they often require large quantities of labeled training data, which is frequently not readily available. This is why models that can reliably classify data based on only few examples for each class are an interesting topic of research. One such kind of model are prototypical networks. They utilize few samples to create an embedding space in fewer dimensions, in which similar data points cluster around a single class prototype. In this thesis, we investigate if the embedding space of a prototypical network makes for a good approach for the purpose of visualizing high-dimensional, unstructured data. The goal is to reduce the dimensionality of the data in such a way that the highdimensional relations and structures between data points are preserved, resulting in 2D representations of the data that form coherent class clusters in a scatter plot visualization. This approach is compared with, and evaluated against, other well known supervised and unsupervised dimensionality reduction techniques. Through quantitative experiments relying on statistical measures, as well as a qualitative evaluation of our results, we find that our ProtoNet is capable of producing point embeddings in which the spatial separation of classes is as good or better than the other methods.", pages = "62", address = "Favoritenstrasse 9-11/E193-02, A-1040 Vienna, Austria", school = "Research Unit of Computer Graphics, Institute of Visual Computing and Human-Centered Technology, Faculty of Informatics, TU Wien", keywords = "Dimensionality Reduction, Data Visualization, Prototypical Network, ProtoNet, Class Separation, Scatter Plot", URL = "https://www.cg.tuwien.ac.at/research/publications/2025/stoff-2025-pvu/", }