Tutorial on Comparative Visualization: Interactive Designs and Algorithms Depending on Data and Tasks

Data comparison in various domains can be effectively supported by visual analytics solutions combining interactive visualization and algorithmic analysis. The design of such solutions should match the comparison problem at hand: the input data and the task specification. This requires several choices from algorithm to visual design and interaction. Such design choices also need to consider human perception capabilities. Our tutorial presents how the differences in data and task characterizations influence visual-analytical solution designs. We will first present a conceptual framework, which defines a set of dimensions along which the comparison problem is defined. We then show how this specification influences the comparative solution design both in theory and using real world examples. Our tutorial provides visualization designers with a means to systematize domain problem analysis and to learn which algorithms, visual designs and interactions to use when, also taking into consideration human perception and cognition capabilities. The tutorial is held at a beginner to an intermediate level. 1 TUTORIAL MOTIVATION Comparison is an important data analysis task in many domains such as biology, finance, transportation or medicine. Comparison tasks can be effectively supported by a combination of interactive visualization and algorithmic data analysis [19, 20]. The comparative systems needs to match the comparison problem at hand: input data and the task [4,27,31,35]. The solutions need to suitably combine visual design, interaction and algorithmic analysis, taking into account the limited user’s cognition capabilities. This is difficult, as many choices need to be made when designing comparison solutions. Which algorithms to use? Which similarity function to use? Which visualization type should be used? Which visual encoding should be used? How to support comparison by interaction? Therefore, visualization designers need a comprehensive characterization of comparison problems for developing suitable visual comparison systems. This tutorial addresses this need and offers a detailed description of comparison problems (data and tasks) together with their implications for the comparison solutions. We will present both theoretic guidelines and practical examples of best practices. Our tutorial builds upon visualization literature on task and data characterizations, comparison visualization techniques such as [1, 4, 15, 19–21, 24, 25, 27, 35, 38] as well as our extensive experience with designing visual analytics solutions for comparative tasks across various application domains such as finance [45], medicine [42], biology [13, 23, 28], transportation [44], meteorology [43], and perception studies [9, 46]. 2 TUTORIAL ORGANIZATION The tutorial is organized along the visual comparison process. It starts with the comparison problem description as a combination of input data and task. Then, comparison operations turn input data into outputs. This operation can be done algorithmically, ∗e-mail: office@gris.tu-darmstadt.de †e-mail: hjschulz@cs.au.dk ‡e-mail: N.Kerracher@napier.ac.uk §e-mail: margit@igw.tuwien.ac.at visually or in a combined way. The outputs are shown to the user in an interactive visualization. The visual design should include interaction and should consider user’s cognitive capabilities. The planned schedule follows this overall outline: (Introduction – ca. 10-15 minutes) Part 1: Specification of Comparison Problems: Data and Tasks We will present a set of conceptual dimensions along which the comparison data and task can be characterized. (ca. 30 minutes, N. Kerracher + T. von Landesberger) Part 2: Algorithmic Comparison: We will present computational methods for comparing and relating data items ranging from matching approaches, via classification methods, to clustering algorithms. (ca. 45 minutes, H.-J. Schulz) (Coffee Break – ca. 30 minutes) Part 3: Visual Design and Interaction: We will present types and real examples of visual designs and interaction techniques fitting to various comparison problem specifications (ca. 45–60 minutes, T. von Landesberger) Part 4: Perception and Cognition: We will present the perceptual and cognitive mechanisms relevant for visual comparison, and underline them with insights and guidelines derived from studies on visual comparison. (ca. 30 minutes, M. Pohl + K. Ballweg) (Wrap-Up and Q&A – ca. 10–15 minutes) 2.1 Part 1: Comparison Problems: Data and Tasks Data and task characterization is needed to describe the comparison problem to be solved. Proper data and task characterization are a prerequisite for finding suitable visual comparison solutions. For example, financial analysts need to analyze the implications of a financial crisis by determining the change of financial institutions and their connections in a financial network after the crisis, to the reference network before the crisis, e.g., 2010 to 2005. Therefore, we will first present important dimensions for specifying comparison problems that can be used across application domains. We will include many examples of specifications along these dimensions. In the financial example, these are: 1) the comparison purpose (i.e., what is the target output) and 2) the comparison input, on which the task is performed [4]. The comparison purpose is to determine change. This change is determined by comparing the existence of nodes and edges in the network of 2010 to the network of 2005. Thus, the comparison input consists of a) what is compared – i.e., nodes and edges of the two networks; b) what is compared to what – i.e., 2010 and 2005 networks; and c) according to which aspect – i.e., their existence. We will explain each dimension in detail both conceptually and using real examples. This set of dimensions will then be used in the next sections for characterizing suitable solutions. 2.2 Part 2: Algorithmic Comparison Data comparison is often supported by algorithmic means. In these cases, the visualization’s responsibility is not so much to allow for visual comparison, but for the interactive steering of the used algorithm and for the interactive exploration of its result. We can subdivide algorithmic comparison approaches by the cardinality of the involved data. 1-to-1 Comparison: Matching aims to map one set of data onto another set of data as closely as possible. Common instances are graph matching, pattern matching, and string matching – both exact and inexact – that have applications from de-duplication of data records [12] to detecting errors in crowd-sourced data [22]. 1-to-Many Comparison: Classification is the process of assigning one data item a suitable class out of a finite number of predefined classes. Common examples are deciding for a pattern of incoming network connections whether these are malicious or not, and part of speech tagging of words in a sentence. Algorithmic means to do that are decision trees [40] and the k-nearest neighbor algorithm [36], for example. Many-to-Many Comparison: Clustering compares all data items to all others to establish groups of similar data items. Clustering can be done bottom-up by successively agglomerating data items into larger groups, or top-down by successively partitioning a single large group that encompasses all data items. Clustering is used in a variety of application domains from biomedicine [26] to the analysis of time-oriented data [5]. This part will give a comprehensive overview of and introduction to these three flavors of algorithmic comparison, explain their commonalities and differences, their respective uses and outcomes, as well as potential difficulties and pitfalls in their use. 2.3 Part 3: Interactive Visualization Design This section will present the implications of comparison problem specification (see Part 1) on the visual-interactive design of comparative visualizations both as means of data comparison and as showing comparison result of algorithmic calculation (see Part 2). There are many comparative visualizations, see surveys [2–4,10, 21,24,25,30,41]. The main categories independent of the input data are [19, 20]: juxtaposition (showing compared data in separate coordinate space), superposition (overlaying the data in the same coordinate space), and explicit representation of relationships. These basic representations are often combined or have variants such as superposition in space (e.g., overlay) or superposition in time (e.g., animation, interchangeable display). These categories are, however, not discriminative enough for providing detailed design guidelines. For example, juxtaposition has several variants, even for one data type and often require mixed types such as juxtaposition with explicit encoding. We will present more detailed discrimination of the designs according to the specification of the comparison purpose and of the comparison inputs. Moreover, interaction supports comparative visualizations, e.g., by bringing closer the items to be compared. Specific interaction techniques have been proposed e.g., [32, 39]. The choice of the interaction technique should support the comparison goal. We will present examples of such interaction techniques. We will first present the main categories of comparative visualizations and their concrete design specifications as implications of data and task characterization. We will present real examples of the techniques as well as discuss their advantages and disadvantages. This will provide the audience guidelines for choosing suitable visual design for the task at hand. 2.4 Part 4: Perception and Cognition Respecting perception and cognition for visual data comparison is important because multiple studies have shown that the human similarity perception strongly diverges from mathematical measures [11, 17, 33, 34]. Furthermore, studies have shown that there are factors tricking the human perception and cognition to over/underestimate the data items similarity [8]. We will first present relevant