Bastian Rieck

Multivariate Data Analysis Using Persistence-Based Filtering and Topological Signatures

The extraction of significant structures in arbitrary high-dimensional data sets is a challenging task. Moreover, classifying data points as noise in order to reduce a data set bears special relevance for many application domains. Standard methods such as clustering serve to reduce problem complexity by providing the user with classes of similar entities. However, they usually do not highlight relations between different entities and require a stopping criterion, e.g. the number of clusters to be detected. In this paper, we present a visualization pipeline based on recent advancements in algebraic topology. More precisely, we employ methods from persistent homology that enable topological data analysis on high-dimensional data sets. Our pipeline inherently copes with noisy data and data sets of arbitrary dimensions. It extracts central structures of a data set in a hierarchical manner by using a persistence-based filtering algorithm that is theoretically well-founded. We furthermore introduce persistence rings, a novel visualization technique for a class of topological features-the persistence intervals-of large data sets. Persistence rings provide a unique topological signature of a data set, which helps in recognizing similarities. In addition, we provide interactive visualization techniques that assist the user in evaluating the parameter space of our method in order to extract relevant structures. We describe and evaluate our analysis pipeline by means of two very distinct classes of data sets: First, a class of synthetic data sets containing topological objects is employed to highlight the interaction capabilities of our method. Second, in order to affirm the utility of our technique, we analyse a class of high-dimensional real-world data sets arising from current research in cultural heritage.

Persistent homology for the evaluation of dimensionality reduction schemes

High‐dimensional data sets are a prevalent occurrence in many application domains. This data is commonly visualized using dimensionality reduction (DR) methods. DR methods provide e.g. a two‐dimensional embedding of the abstract data that retains relevant high‐dimensional characteristics such as local distances between data points. Since the amount of DR algorithms from which users may choose is steadily increasing, assessing their quality becomes more and more important. We present a novel technique to quantify and compare the quality of DR algorithms that is based on persistent homology. An inherent beneficial property of persistent homology is its robustness against noise which makes it well suited for real world data. Our pipeline informs about the best DR technique for a given data set and chosen metric (e.g. preservation of local distances) and provides knowledge about the local quality of an embedding, thereby helping users understand the shortcomings of the selected DR method. The utility of our method is demonstrated using application data from multiple domains and a variety of commonly used DR methods.

Structural Analysis of Multivariate Point Clouds Using Simplicial Chains

Topological and geometrical methods constitute common tools for the analysis of high‐dimensional scientific data sets. Geometrical methods such as projection algorithms focus on preserving distances in the data set. Topological methods such as contour trees, by contrast, focus on preserving structural and connectivity information. By combining both types of methods, we want to benefit from their individual advantages. To this end, we describe an algorithm that uses persistent homology to analyse the topology of a data set. Persistent homology identifies high‐dimensional holes in data sets, describing them as simplicial chains. We localize these chains using geometrical information of the data set, which we obtain from geodesic distances on a neighbourhood graph. The localized chains describe the structure of point clouds. We represent them using an interactive graph, in which each node describes a single chain and its geometrical properties. This graph yields a more intuitive understanding of multivariate point clouds and simplifies comparisons of time‐varying data. Our method focuses on detecting and analysing inhomogeneous regions, i.e. holes, in a data set because these regions characterize data in a different manner, thereby leading to new insights. We demonstrate the potential of our method on data sets from particle physics, political science and meteorology.

Exploring and comparing clusterings of multivariate data sets using persistent homology

Clustering algorithms support exploratory data analysis by grouping inputs that share similar features. Especially the clustering of unlabelled data is said to be a fiendishly difficult problem, because users not only have to choose a suitable clustering algorithm but also a suitable number of clusters. The known issues of existing clustering validity measures comprise instabilities in the presence of noise and restrictive assumptions about cluster shapes. In addition, they cannot evaluate individual clusters locally. We present a new measure for assessing and comparing different clusterings both on a global and on a local level. Our measure is based on the topological method of persistent homology, which is stable and unbiased towards cluster shapes. Based on our measure, we also describe a new visualization that displays similarities between different clusterings (using a global graph view) and supports their comparison on the individual cluster level (using a local glyph view). We demonstrate how our visualization helps detect different—but equally valid—clusterings of data sets from multiple application domains.

UNWRAPPING HIGHLY-DETAILED 3D MESHES OF ROTATIONALLY SYMMETRIC MAN-MADE OBJECTS

Rotationally symmetric objects commonly occur at archaeological finds. Instead of creating 2D images for documentation purposes by manual drawing or photographic methods, we propose a method based on digitally colored surface models that are acquired by 3D scanners, thereby including color information. We then transform these highly-detailed meshes using simple geometrical objects such as cones and spheres and unwrap the objects onto a plane. Our method can handle curved vessel profiles by dividing the surface into multiple segments and approximating each segment with a cone frustum that serves as an auxiliary surface. In order to minimize distortions, we introduce a simple quality measure based on distances of points to a fitted cone. We then extend our method to approximately spherical objects by fitting a sphere on the surface of the object and applying a map projection, namely the equirectangular projection known from cartography. Our implementation generates true-to-scale images from triangular meshes. Exemplary results demonstrate our methods on real objects, ranging from small and medium-sized objects such as clay cones from the Ancient Orient and figural friezes of Greek vessels to extremely large objects such as the remains of a cylindrical tower of Heidelberg Castle.

Clique Community Persistence: A Topological Visual Analysis Approach for Complex Networks

Complex networks require effective tools and visualizations for their analysis and comparison. Clique communities have been recognized as a powerful concept for describing cohesive structures in networks. We propose an approach that extends the computation of clique communities by considering persistent homology, a topological paradigm originally introduced to characterize and compare the global structure of shapes. Our persistence-based algorithm is able to detect clique communities and to keep track of their evolution according to different edge weight thresholds. We use this information to define comparison metrics and a new centrality measure, both reflecting the relevance of the clique communities inherent to the network. Moreover, we propose an interactive visualization tool based on nested graphs that is capable of compactly representing the evolving relationships between communities for different thresholds and clique degrees. We demonstrate the effectiveness of our approach on various network types.

Visualization of 4D Vector Field Topology

In this paper, we present an approach to the topological analysis of four‐dimensional vector fields. In analogy to traditional 2D and 3D vector field topology, we provide a classification and visual representation of critical points, together with a technique for extracting their invariant manifolds. For effective exploration of the resulting four‐dimensional structures, we present a 4D camera that provides concise representation by exploiting projection degeneracies, and a 4D clipping approach that avoids self‐intersection in the 3D projection. We exemplify the properties and the utility of our approach using specific synthetic cases.

Persistent Homology in Multivariate Data Visualization

Technological advances of recent years have changed the way research is done. When describing complex phenomena, it is now possible to measure and model a myriad of different aspects pertaining to them. This increasing number of variables, however, poses significant challenges for the visual analysis and interpretation of such multivariate data. Yet, the effective visualization of structures in multivariate data is of paramount importance for building models, forming hypotheses, and understanding intrinsic properties of the underlying phenomena. This thesis provides novel visualization techniques that advance the field of multivariate visual data analysis by helping represent and comprehend the structure of high-dimensional data. In contrast to approaches that focus on visualizing multivariate data directly or by means of their geometrical features, the methods developed in this thesis focus on their topological properties. More precisely, these methods provide structural descriptions that are driven by persistent homology, a technique from the emerging field of computational topology. Such descriptions are developed in two separate parts of this thesis. The first part deals with the qualitative visualization of topological features in multivariate data. It presents novel visualization methods that directly depict topological information, thus permitting the comparison of structural features in a qualitative manner. The techniques described in this part serve as low-dimensional representations that make the otherwise high-dimensional topological features accessible. We show how to integrate them into data analysis workflows based on clustering in order to obtain more information about the underlying data. The efficacy of such combined workflows is demonstrated by analysing complex multivariate data sets from cultural heritage and political science, for example, whose structures are hidden to common visualization techniques. The second part of this thesis is concerned with the quantitative visualization of topological features. It describes novel methods that measure different aspects of multivariate data in order to provide quantifiable information about them. Here, the topological characteristics serve as a feature descriptor. Using these descriptors, the visualization techniques in this part focus on augmenting and improving existing data analysis processes. Among others, they deal with the visualization of high-dimensional regression models, the visualization of errors in embeddings of multivariate data, as well as the assessment and visualization of the results of different clustering algorithms. All the methods presented in this thesis are evaluated and analysed on different data sets in order to show their robustness. This thesis demonstrates that the combination of geometrical and topological methods may support, complement, and surpass existing approaches for multivariate visual data analysis.

Interactive Similarity Analysis and Error Detection in Large Tree Collections

Automatic feature tracking is widely used for the analysis of time-dependent data. If the features exhibit splitting behavior, it is best characterized by tree-like tracks. For a large number of time steps, each with numerous features, these data become increasingly difficult to analyze. In this paper, we focus on the problem of comparing and contrasting hundreds to thousands of trees to support developmental biologists in their study of cell division patterns in embryos. To this end, we propose a new visual analytics method called structure map. This two-dimensional, color-coded map arranges trees into tiles along a Hilbert curve, preserving a tree similarity measure, which we define via graph Laplacians. The structure map supports both global and local analysis based on user-selected tree descriptors. It helps analysts identify similar trees, observe clustering and sizes of clusters within the forest, and detect outliers in a compact and uniform representation. We apply the structure map for analyzing 3D cell tracking from two periods of zebrafish embryogenesis: blastulation to early epiboly and tailbud extension. In both cases, we show how the structure map supported biologists to find systematic differences in the data set as well as detect erroneous cell behaviors.

Agreement Analysis of Quality Measures for Dimensionality Reduction

High-dimensional data sets commonly occur in various application domains. They are often analysed using dimensionality reduction methods, such as principal component analysis or multidimensional scaling. To determine the reliability of a particular embedding of a data set, users need to analyse its quality. For this purpose, the literature knows numerous quality measures. Most of these measures concentrate on a single aspect, such as the preservation of relative distances, while others aim to balance multiple aspects, such as intrusions and extrusions in k-neighbourhoods. Faced with multiple quality measures with different ranges and different value distributions, it is challenging to decide which aspects of a data set are preserved best by an embedding. We propose an algorithm based on persistent homology that permits the comparative analysis of different quality measures on a given embedding, regardless of their ranges. Our method ranks quality measures and provides local feedback about which aspects of a data set are preserved by an embedding in certain areas. We demonstrate the use of our technique by analysing quality measures on different embeddings of synthetic and real-world data sets.