DataChopin Designing Interactions for Visualisation Composition in a Co-located, Cooperative Environment

. This article presents our interaction design for DataChopin, based on an extensive survey and classi(cid:28)cation of visualisation for exploratory data analysis. Its distinctive characteristics are the use of a large-scale display wall as a shared desktop, as well as (cid:29)exible composition mechanisms for incremental and piece-wise construction of visualisations. We chose composability as a guiding principle in our design, since it is essential to open-ended exploration, as well as collaborative analysis. For one, it enables truly exploratory inquiry by letting users freely examine di(cid:27)erent combinations of data, rather than o(cid:27)ering a predetermined set of choices. Perhaps more importantly, it provides a foundation for data analysis through collaborative interaction with visualisations. If data and visualisations are composable, they can split into independent parts and recombined during the analytical process, allowing analysts to seamlessly transition between closely-and loosely-coupled work.


Introduction
The need for better understanding of abstract data is not new, although it is exacerbated by the growing ease and speed of data acquisition. When automatic methods fail, human background knowledge and intuition are required. Yet as Norman [13] points out, human cognitive abilities are highly constrained and our real ingenuity lies in the ability to devise external aids that enhance them. While our cognitive capabilities for storing and manipulating data may be limited, we have evolved to perform many analytical processing tasks visually. Therefore, a common approach is to devise visual representations of data. However, the choice of representation is not trivial and depends heavily on the task at hand. Consequently, a wide range of visualisation systems have been developed, designed specically for certain tasks. While such systems with predened components are typically eective in their intended area of application, they are often too limited for open-ended, exploratory analysis, which requires the ability to manipulate and tailor representations based on emerging questions and insights. In order to address this, researchers have systematically studied the structure of graphical representations, along with the rules by which visualisations are constructed. Our proposed system continues this line of research, building on existing theoretical and formal models to arrive at a practical implementation and explore suitable interaction techniques and metaphors for cooperative visualisation specication.
A key promise of novel interaction technologies are better ways for people to work cooperatively. Concepts that were previously only explored in research such as large-scale environments augmented with interactive capabilities are becoming technologically feasible. We aim to utilise such technologies to create engaging experiences that put multiple analysts in a shared environment and elicit contextual knowledge from these analysts. The importance of contextual knowledge for the analytical process was already pointed out by Cleveland: Conclusions spring from data when this information is combined with the prior knowledge of the subject under investigation. [6, p. 5] When multiple analysts join eorts, and are given eective tools to share and communicate their visions, the potential for more diverse and unexpected insights stands to grow accordingly.

Related Work
As part of this review, we examine existing systems that allow exible visualisation specication for exploratory data analysis. Since a multitude of such systems have been developed, we do not attempt to create an exhaustive survey.
Instead, we aim to highlight conceptual dierences based on notable examples. Furthermore, we identify and categorize the predominant interaction methods, highlighting their commonalities and dierences. Since our goal is the design of an interface for ad-hoc end-user composition and collaboration, we place particular focus on how the dierent conceptual models relate to interface mechanisms and metaphors for visualisation construction.
In order to develop exible systems for visualisation specication it is necessary to understand the integral components and structure of graphics. The systematic analysis graphical representations, lies at the core of an important set of visualisation theories, sometimes referred to as structural theories of graphics. Much of this research can be traced back to the Semiology of Graphics by Bertin [2], which represents one of the rst attempts to interpret graphics as a language with formal rules. Another prevalent theoretical model is that of a visualisation pipeline, as popularised by Card et al. [5]. This model provides a description of the visualisation process as a sequence of transformation steps.
Depending on the chosen theoretical perspective, certain paradigms for composing visualisations naturally lend themselves. Perhaps the most obvious form of visualisation specication is an imperative algorithm that issues drawing commands to produce graphical primitives. Such specications provide ne grained control, but require familiarity with programming concepts, such as variables and control ow. However, the approach as its strengths, as evidenced by the popularity of Processing [7]. In contrast to imperative algorithms, more recent approaches employ a declarative paradigm, placing emphasis on what to display rather than how to produce it. Such a specication features a description of a graphical scene, and allows connecting data attributes to visual attributes of graphical elements. Libraries based on the data binding model are Protovis [3] and its successor D3 [4]. On another end of the spectrum, researchers have created automated presentation tools eliminate the need for a specication entirely.
Such research advanced the study of graphic primitives and pioneered key ideas, such as composition algebras for graphical marks [11,14]. Extending further on the ideas of structural theories of graphics, researchers have developed sophisticated graphics grammars [21,20], resulting in specication languages that allow the assembly of statistical graphics from ne-grained, modular units of composition. These ideas also provide the foundation for Vega [15], the low-level declarative language behind Lyra as well as VizQL [8], the query-based language behind Tableau. Meanwhile, the prominence of the visualisation pipeline model has lead to the adoption of a data ow paradigm in many systems. The pipeline structure provides a blueprint for the implementation of recongurable visualisation components, such as those in VTK [16].
In addition to underlying conceptual models, there are dierent options for exposing them through graphical interfaces. These aspects are closely related, Wilkinson, who claim that choosing from a limited charts gives the user an impression of having explored data rather than the experience [21, p. 2]. This appears problematic especially for exploratory analysis, as it oers no way for users to produce visual representations beyond those explicitly supported by the system.

Text and Preview A very basic form of graphical support is an environment
where text-based specication is accompanied by a preview window. This approach is adopted by software like Processing Development Environment [7] and GPL [21]. The former employs an imperative programming style, whereas the latter is a proprietary implementation of Wilkinson's graphics grammar. A modied version of the grammar is publicly available in the form of the ggplot2 [20] module for the R statistical computing environment. While text input is often challenging on touch interfaces, it can be assisted through auto-completion or direct manipulation of the parse tree.
Tokens and Slots More sophisticated graphical interfaces allow data binding via property sheets, or by dragging and dropping tokens on designated regions of the interface. Such a model is realised in software like ILOG Discovery [1], Polaris [17], and its successor Tableau [8]. These interfaces often incorporate abstract representations of the variables in a data set, which can be manipulated via drag-and-drop gestures. This commonly represents data binding, whereby variables are bound to properties of the visual representation. Furthermore, Polaris and Tableau also inherit ideas from visualisation grammars, such as a compositional algebra to specify combinations and nestings of data variables.
Boxes and Wires Another interaction concept are boxes and wires, also known as the node-link model of visual programming, which is a natural t for the data ow model of visualisation. This approach combines a visual notation with expressive power of text-based specication languages. Complex ows are created by placing processing units that act as operators on the data, and subsequently connecting their inputs and outputs to create a graph. Such a model employed by VTK [16] in the form of VTK Designer.
Pipeline Stages Such interfaces are characterised by high-level abstractions focussed on the application domain. For example, they might be restricted to pipelines with a limited set of stages, which are directly represented within the interface as text or icons. This style is followed by the Lark [18] application, which uses a pipeline with customisation points at three stages: analytical abstraction, spatial layout, and presentation. Furthermore, LIVE Singapore! Data Browser [10] and Datacollider from MIT's SENSEable City Lab apply similar models. Outside of the visualisation domain, other notable interfaces using very specialised programming models are Reactable [9] for musical composition and Kodu [12] for specifying simple behaviours in games.
Drawing Canvas Generally, the interaction model of graphical applications is not a good t for the task of visualisation specication, as manual manipulation of marks quickly becomes becomes repetitive and tedious. However, paired with facilities for automation and data binding, this interaction style can become feasible. In this respect, the web-based Lyra [15] application is worth mentioning, as it is inspired by Victor's interface for drawing dynamic visualisations [19]. The users create and arrange visual marks on the canvas through direct manipulation.
Subsequently, data variables are dragged onto various anchors in order to bind data to visual properties.

Design Process
Based on our review, we assessed the identied interaction concepts for use colocated, collaborative settings. Ultimately, this process informed the design our nal artefact, named DataChopin. The system was designed from the ground up for co-located, multi-user interactions. Therefore, data sets and visualisations are associated with user accounts. Once a user authenticates with the system, their presence is indicated by a top-level menu element on the shared desktop, which features an avatar and provides access to personal content. Data Interactions The interaction metaphor for selecting and manipulating data attributes was inspired by poker chips. Our design intuition was that these elements would introduce interesting dynamics to a collaborative analysis process.
Poker chips are employed in a variety of tabletop games, enabling playful mechanics and social interactions. They carry associations with collecting, exchanging, and negotiating. Our goal was to capture the aordances of poker chips, while enhancing their digital counterparts with capabilities for data analysis. Another conguration that we explored was a single, shared drawing canvas that multiple users gathered around, as illustrated by gure 2a. This congurations supports a closely coupled style of cooperation, and is either based on a single data of shared interest, or on multiple data sets from the repositories of dierent users, resulting in a combined, layered visualisation. The latter is made possible due to composability being an integral feature of the system. Finally, as a single view can lead to contention, we also experimented with compromises between a single, shared and multiple, independent views. One conguration that appears promising are two drawing canvases, with a number of shared data repositories in between them, depicted in gure 2b. That way, two groups can work independently, both having access to the same data sets. If the data used by both groups comes from the same data sets, the visualisations are always compatible, meaning that the groups can interchange and combine parts of their visualisation specications at any given time.

Discussion
Our work lead us to survey the spectrum of conceptual models and interaction concepts for visualisation specication, weighing their associated trade-os in the specic context of exploratory analysis in collaborative environments. Early on, we ruled out static chart typologies, due to concerns that their rigidity would stie creativity. Furthermore, piece-wise and iterative specication is considered benecial for open-ended analysis, and forms an important cornerstone for collaboration in our proposed system. Text-based specications have proven eective, especially for seasoned users who are familiar with the syntax and semantics of the underlying specication language. However, they are in conict with direct manipulation principles and pose challenges with regard to text input on touch-based interfaces. While boxes and wires provide a visual notation capable of modelling general-purpose programming languages, their generality comes at the cost of usability. In our experience, we found the domain of exploratory analysis suciently constrained to employ a special-purpose abstraction. In our classication, DataChopin is a hybrid of tokens and slots, combined with a visualisation canvas for declarative data binding.
So far, we have conducted informal evaluations our system, and our initial experiences have been positive. Cooperating in a co-located setting successfully elicited discussions about the data and participants were quick to share their interpretations. The use of a multi-user, shared-desktop environment was commonly regarded as benecial. In contrast to the single-user, personal systems that participants were accustomed to, the idea of multiple analysts working in tandem was perceived as empowering. Rather than a single person being in charge and driving the analytical process, the interface enabled them to perform actions in parallel and pursue smaller tasks independently. Therefore, we continue to focus our eorts on placing participants in shared interaction environments, aiming to leverage the implicit and explicit communication channels to stimulate creativity and assist analysis. In future work, we are planning more formal evaluations to assess the expressiveness and eectiveness of the proposed compositional model.

Conclusion
Our review has shown that HCI research on cooperative visualisation specication is still lacking. While some systems support distributed, asynchronous collaboration, few focus on co-located, synchronous settings. With the exception of Lark, the majority of existing interfaces were designed for single-user, personal environments. This article represents another step towards closing the research gap. We have classied predominant interaction methods for visualisation specication, and derived a design specically aimed at facilitating cooperation in a shared interaction environment. The result is DataChopin, a system for largescale, shared-desktop environments, based on the premise of composable visualisations. Often, formal visualisation models have been studied in theory and divorced from HCI considerations. In contrast to that, our work presents a practical approach, covering the design and implementation of a working prototype.