Effective temporal graph layout: A comparative
Transcription
Effective temporal graph layout: A comparative
Effective temporal graph layout: A comparative study of animation versus static display methods Michael Farrugiaa, ∗ and Aaron Quigleyb a University College Dublin, 8 Talbot Down, Dublin15, Ireland. E-mail: [email protected] b Chair of Human Computer Interaction, School of Computer Science, The University of St. Andrews, UK. E-mail: [email protected] ∗ Corresponding author. Abstract Graph drawing algorithms have classically addressed the layout of static graphs. However, the need to draw evolving or dynamic graphs has brought into question many of the assumptions, conventions and layout methods designed to date. For example, social scientists studying evolving social networks have created a demand for visual representations of graphs changing over time. Two common approaches to represent temporal information in graphs include animation of the network and use of static snapshots of the network at different points in time. Here, we report on two experiments, one in a laboratory environment and another using an asynchronous remote web-based platform, Mechanical Turk, to compare the efficiency of animated displays versus static displays. Four tasks are studied with each visual representation, where two characterise overview level information presentation, and two characterise micro level analytical tasks. For the tasks studied in these experiments and within the limits of the experimental system, the results of this study indicate that static representations are generally more effective particularly in terms of time performance, when compared to fully animated movie representations of dynamic networks. Information Visualization (2011) 00, 1 -- 18. doi:10.1057/ivs.2010.10 Keywords: graph drawing; dynamic social networks; information visualisation; evaluation; animation Introduction Received: 27 August 2009 Revised: 23 August 2010 Accepted: 13 September 2010 Real-world networks are not static, they are in a state of constant evolution. Networks, such as social science networks, biological networks and computer networks, are constantly changing with new nodes added and new relations forming, while old relations may either persist or end. In this article, we focus on dynamic social networks and their effective visualisation. Despite the lack of widespread research publications on dynamic social networks, Doreian and Stokman1 report that research on dynamic networks is increasing. Recently, Volume 325 (July 2009) of the Science Journal featured a special edition on Complex Systems and Networks. In this feature, written by the most prominent authors in network science, the study of dynamic networks was identified as one of the current main challenges in network theory. It is recognised that collecting longitudinal data is a time consuming and challenging task. This explains why network data is typically only collected once for analysis. Owing to this, the field of social network analysis has often concentrated on single snapshot instances of a network. Thanks in part to the proliferation of online technologies and social media, the data collection bottleneck has been partially overcome and dynamic network data sources have become more accessible. © SAGE Publications, 2011. 1473-8716 Information Visualization Vol. 00, 0, 1 – 18 Farrugia and Quigley Dynamic social network analysis on longitudinal data can be useful for investigating long-term human behaviour. Christakis and Fowler2 studied the dynamics of smoking behaviour over a period of 32 years. Smoking decreased throughout the data collection period and interestingly connected groups of smokers stopped smoking together suggesting a network effect. In another study, van Duijn et al,3 studied the evolution of friendships in University freshmen. Their results show that physical proximity of actors and the individual similarities between actors, are only contributors to the beginning of friendship formation. Dynamic network analysis can also be important in criminal investigations. Xu et al4 describe the processes and measures applicable when using dynamic network analysis in a criminal context and also make effective use of visualisation, including animation, to present the evolution process of criminal networks. From the earliest research in social network analysis, visualisation has been a central tool. Moreno, the inventor of sociometry,5 devised the first graphical representation of a relationship between two entities and called it the sociogram. Since Moreno’s days there has been a significant increase in complexity and improvement in the methods for how social networks are drawn.6 However in network visualisation, the temporal aspects of the data within the network have often been ‘flattened out’ or ignored. The realisation that the study of different versions of a network over time can lead to insights not possible through static display alone has given rise to an increasing interest in dynamic graphs within the visualisation community. A testament to this, is the fact that at least three major conference competitions, GD’98,7 InfoVis’048 and VAST’089 contained tasks requiring the visualisation and characterisation of dynamic networks. Beyond the problem of developing either a graph layout technique or a new visual design for dynamic networks lies the presentation of the visualisation. Classically, there are three ways in which dynamic networks, visualised as node-link diagrams, can be presented. The most common presentation of dynamic graphs is to animate the different layouts for each time point, and produce a movie of the dynamic network.6,10–13 Alternatively, the individual layouts can be displayed adjacent to each other in one screen akin to a photo album with thumbnail images.14,15 In a variant of this presentation, the node link diagrams can be displayed as static images in sequence, similar to a slideshow.16,17 When it comes to evaluating the mode of presentation of dynamic networks, studies usually either rely on the general intuition that animation is beneficial, or else ask for experts to judge the usefulness of animation.4 In reality, no empirical studies have been conducted on the actual benefits of animation when visualising dynamic social networks. In this study, we attempt to empirically evaluate the benefits of animation and understand the tasks and cases for which animation is beneficial over static images. 2 As part of our evaluation, we conduct two experiments using different experimentation platforms. The first experiment is a laboratory user study with a group of computer science students. In the second experiment, we attempt to generalise the results obtained in the laboratory by extending the study to anonymous users, who might not have experience with social network visualisations, in a remote experiment using Mechanical Turk (MTurk). As part of our results, and discussion we detail our experiences using a remote experimentation platform and the implications it has on visualisation experiments. The rest of the article is organised as follows. In Section ‘Related work’, we review related work in the area of dynamic social network visualisation, animation and visual perception, including approaches to visualisation evaluation with an emphasis on remote experiments for visualisation. In Section ‘Experiment setup and materials’, we describe our experiments’ design, materials and methodology used to compare two visual representations. In Sections ‘Results’ and ‘Discussion’, we report on the results and then discuss them. Finally, we derive conclusions based on the results and discuss areas for further work, both in the evaluation and implementation of such visual systems. Related Work This work draws on three principal areas of visualisation research; graph drawing, visual perception and visualisation evaluation methods. Dynamic graph drawing provides the technical background to visualise networks following desirable criteria of graph layouts with a temporal component. As the aim of the study is to compare animated and static modes of presentation, research on the efficiency of both mediums as a perceptual aid in different visualisation applications, is reviewed. Part of this research includes a user study, where a relatively new platform that takes advantage of a web-based micro-market to engage a user study population. Early studies this platform, report promising results if the necessary care is take to interpret the results, in the appropriate context. Recent work on the applicability of this platform for the purpose of resourcing participants and conducting online experiments is also reviewed. Dynamic graph drawing In its most basic form a graph consists of a number of nodes, representing elements of information and the interrelationships between these elements, or edges. Graph drawing is the challenge of determining a geometric position for every node and a route for every edge, so that a picture of the graph can be rendered. Dynamic graph drawing is an area within graph drawing that studies the layout of graphs with a temporal component. This dynamic component can come about from new Effective temporal graph layout nodes or edges being added to the network or existing ones being removed or updated. For the purposes of this research, we focus specifically on social networks. Graph theory provides a solid framework onto which the concepts and ideas from such networks can be mapped. This brings a vocabulary which can be employed to denote structural properties and a set of primitive concepts and mathematical operators to measure these structural properties that are representing social structures.18 Given the natural representation of a social network as a graph, graph drawing itself is the natural approach to the visualisation of social networks. Graph drawing is a well-researched area in computer science and several good references exist on the subject.19–22 In addition, Freeman6 provides a detailed reference on which of these algorithmic approaches are best suited to social network visualisation. In traditional graph drawing, a single static version of the network is drawn. However, when drawing ‘dynamic networks’ several related versions of the same basic network need to be drawn. Here, the time dimension adds an additional constraint beyond the traditional optimisation of aesthetic criteria20,23,24 with static graphs. When drawing temporal dynamic graphs it is generally considered beneficial to keep node movement between time periods at a minimum. This is to preserve the mental map between different related layouts.25,26 Where dynamic social network visualisation has been studied, the focus has been predominantly on the formation of network structures. The most common approach to the dynamic graph drawing problem builds upon the flexible nature of force directed layout algorithm27 to control node movement. These layout algorithms28,29 often include a stiffness parameter to dampen node movement between successive time-steps to help maintain a smooth change to the layout. When the sequence of graph versions are known before the drawing, the complete layout information can be exploited to improve the node placement. These class of graph drawing techniques are called offline layout algorithms. In one of these techniques, Diehl and Görg30 use the supergraph of all graph iterations to calculate the node positions. These positions are then used to inform the individual node locations at the drawing of each time step. Lee et al31 extend layout heuristics first applied to static graphs, that of simulated annealing by Davidson and Harel.32 They formulate the dynamic graph drawing question as an optimisation problem, whereby they attempt to minimise a cost function of graph changes between iterations. Although encoding different graph criteria separately gives the user more control over the layout, these algorithms are typically quite slow to execute. Dwyer and Gallagher33 employ a two and a half-dimensional layout where time is mapped to the third dimension providing a 3D layered view of the changes in the data over time. In this study, we fix the parameters related to graph layout. We employ the Social Network Image Animator (SoNIA)34 software framework whose focus is on the optimisation of dynamic graph layouts and use it to generate the graph images and videos for our studies. Our approach is to vary the mode of presentation to study the impact on graph understanding. When evaluating dynamic graph drawing algorithms one needs to compare the difference between layouts across different time steps and measure the change to the layout over time.35 Friedrich and Eades36 provide a comprehensible list of criteria and measures of a good animation. Frishman and Tal29 use the distance between nodes in successive layouts to calculate the average displacement of the nodes. They complement this with measures of the energy (stress) of the graph at successive time points as a measure of the ‘niceness’ of the layout. While several good attempts have been made to understand the impact of graph layouts and presentation on the graph comprehension,23,24,37 work on this is still in its formative stages. It is difficult, if not impossible, to claim that a group of measures can be used as an absolute indicator of optimal layout or presentation. Instead, quantitative claims from a new algorithm provide an indication that it improves upon certain criteria, that are currently believed to impact upon perception. Mode of presentation Animation is essentially a sequence of images displayed in rapid succession to give the illusion of movement. Using animation, one can embed time within the network structure itself without changing any additional attributes. Animation is a natural way of presenting temporal information,38 therefore it is hardly surprising that animation is the most popular manner of presenting dynamic networks. One of the most concrete applications of dynamic network visualisation is the study by Moody et al17 that uses dynamic network visualisation to characterise two real and one synthetic dynamic network. Both animation and static images are used to present the results. The animation presentation is used to create movies of the networks, whereas static images are presented using flipbooks (slideshows). In the animations, nodes are free and can change position between different time periods, however in flipbooks all the nodes are fixed between time periods. In flipbooks only one common graph layout is used, an approach that the authors claim is appropriate for relatively sparse networks. The authors also distinguish between continuous and discrete time in networks. Continuous network data is made up of streaming data with known start and end times for each relationship, for example, a network extracted from a log of phone calls. Discrete networks, on the other hand, collect data at specific time intervals. Two classroom interactions from the McFarland classroom network39 are used as an example of a continuous time network. This data consists of a stream of interactions of conversion turns in classrooms. 3 Farrugia and Quigley In their article, the patterns identified in the three networks are presented convincingly using animation. It is clear that animation was the preferred presentation method for analysis and presentation. Perhaps, one reason for this is the fact that animation provided less interactive burden owing to the number of time periods in the networks (for instance, 200 iterations for the syntactic network). Animation has the ability to condense many images in a short space of time and to be executed with one play action. To visualise this data with flipbooks one has to manually browse through all the images in sequence. In addition, the fact that nodes were fixed in flipbooks, does not provide a comparable alternative to animation were the nodes are moving. Animated visualisations of node link diagrams are also popular and successful in visualisation competitions where dynamic network data needs to be analysed. The best two submissions in the Graph Drawing ’98 competition40 were both presented as recorded animations. In Info Vis ’04 the winning submissions analysed the temporal data set of visualisation publications using mainly static network images41 or alternative visualisations to node link diagrams.42 In VAST 2008, there was no single winner but a number of different rewards for significant contributions were awarded. Among these was an award for the best animation which clearly highlighted the structural change occurring in the evolving network.43 The submission explains how the animated visualisations were used to support the analysis and assisted in finding the solution to the problem scenario. Most of the submissions in VAST 2008 used multiple static images of node-link diagrams to show the visualisation. Few of the more advanced systems such as44 added interactive features to allow image scrolling or viewing multiple images in a single window. In addition, some systems extracted the time element from the graph data, and used the time-dimension in a traditional time-based line chart. A notable visualisation in this case was the contribution by the University of Maryland45 that used the Social Action platform46 to visualise the activity of each node throughout the time period as a stacked line chart. Animation studies in node link diagrams Although animation is an intuitive and almost natural way of presenting temporal information, there were very few efforts to evaluate animation as a way of presenting node link diagrams. Most of the evaluations of dynamic networks visualised as node link diagrams study the importance of retaining the mental map in for the graph layout.26 When animation is used for presentation, it is often assumed that animation is the logical and hence only way to represent such information. One of the few studies in this area was conducted by McGrath and Blythe47 who study the impact of motion on viewers’ perception of graph structure. The authors conduct a user study based on an analytical scenario, 4 rather than a purely syntactical task-based experiment, to understand the impact of both motion and layout in understanding the network. Motion is used to facilitate the tracking of the nodes’ between two time periods in the network. The results show that motion has a positive effect when used to understand change in the status of a node. A change to the network layout alone does not have a statistically significant impact unless motion is added to the visualisation. Ware and Bobrow48 investigate motion in node link diagrams, however in this study motion is only used as a highlighting mechanism. In this context, the preattentive property of motion is exploited and used as a visual coding attribute of nodes, such as colour and shape. The three different experiments conducted all suggest that motion is effective as a means of highlighting in node link diagrams. Animation for perception Animation as a mode of presentation, has been studied in different fields with mixed results. Most of this work studies animation in a learning environment.38 In computer science, this was reflected in using animation to explain the internal workings of complex algorithms.49 Similar to the case with dynamic networks, the first developers of algorithm animations believed that animations are beneficial as a learning tool, without empirically evaluating the perceived positive effect. When scientists started to validate this claim by running perception experiments to understand the benefits of animation, animation was not found to perform significantly better or worse than static images.50 Interestingly, followup studies in this area51 attempted to find instances where animation is helpful. In these studies, the benefits of animation were seen when students were allowed to use it in a less rigid environment supplemented with other material. Animation was also found to be useful when the users were allowed to actively interact with the animation by using their own data to be visualised with the animation.52 Inspired by the favourable audience response to Hans Rosling’s presentations at TED 200653 and TED 2007,54 Robertson et al55 addressed the effectiveness of animation in trend visualisation. The authors differentiate animation as used for analysis (when the results are not known), as opposed to when animation as used for the presentation of results (when the results are known). In their study, animation is compared to two static image representations, small multiples and a slideshow view with traces of movement shown. In this study, on multivariate data animation some positive effects of animation are identified in the presentation tasks, yet even for presentation the accuracy in the animated visualisations was lower than the static representations. For the analytical tasks, the two static image representations had better performance in both accuracy and time measures. Interestingly, in the informal survey at the end of the experiment, the Effective temporal graph layout participants thought that animation was easier and more enjoyable than their static counterparts. The appeal of using animation in competition submission40,43 and in explaining network patterns in research publications2,17 can be partially explained by the positive effect animation has when presenting data as described by Robertson. When describing animation as a mode of presentation (in Section ‘Mode of presentation’), we distinguish between ‘a presentation mode’ as opposed to ‘for presentation’. In the context of this article, mode for presentation refers to the way the visualisation is presented, whether it is presented using static images or animation. The graph layout and the node link diagram are the visualisation, whereas animation and static images are two ways that visualisation can be presented. Other work36 has investigated aspects of animation especially the interpolation between keyframes in animations. Yee et al56 explore the case of interpolating between radial-based images and from the informal usability studies find that interpolating the polar coordinates of the nodes radial layout, makes the transition between screens easier to follow. Animation can also be used to assist with transitioning between different visualisations.56–58 When animation is used as a transition the animation is not the main component of the visualisation but animation is used to assist with a shift of viewpoints to make the retention of context or the ‘mental map’ easier. Both the formal evaluation by Heer and Robertson and the informal study by Yee et al report positive results in favour of animation. In this case, animation is used as a tool to assist with transitions rather than as the mode of presentation itself. This application of animation however is quite distinct from animation as a means of presentation, which is the way that animation is being studied in this article. In these experiments, we are investigating animation as a tool in it’s own right. Current systems support for dynamic network visualisation The number of software packages that can be easily used by social scientists to visualise dynamic social networks is limited when compared to the number of packages available for studying static networks. Social scientists who study dynamic networks tend to employ strictly mathematical tools such as the Simulation Investigation for Empirical Network Analysis (SIENA) package,59 which do not support visualisation. If we look at the benefits of information visualisation,60 we can see that dynamic Social Network Analysis can clearly benefit from visualisation. However, a simplistic approach to this problem may further obscure the data and reduce the utility of the visual display with clutter and gratuitous glyph movement. Few visualisation packages support a time element to allow for the network to change over time. For example, Netdraw, which comes packaged with the popular UCINET61 package does not support dynamic networks. Pajek62 supports time-based networks and can generate images of the network at different points in time. Pajek presents the different networks snapshots as single images in a slideshow requiring the user to manually click on a sequence of images. The biggest contribution to date, to the visualisation of dynamic network data, has been from Skye Bender-deMoll and Daniel McFarland,34 who developed an open source software framework called SoNIA to visualise dynamic networks. The main focus of SoNIA is to provide a platform for testing and comparing different graph layout techniques, using dynamic instead of static network data. The network animation in SoNIA is created by joining together a sequence of images of the network, and interpolating the position of nodes between keyframes. The authors define a framework for representing timebased network events. They categorise different time sequences in typical network data sources, and suggest ways how time can be represented in the source data for use in the network visualisations. They introduce the concept of ‘slicing’ event data, as a metaphor to describe a network at a point in time. The slice can be either a thin slice or a thick slice. In a thin slice, the network is extracted for an exact point in time. Thin slices are good to query network data that contains a duration element in the network events. In a thick slice, all the events in particular time window are considered. This is good when network events don’t have duration, or have a small duration, and a time window is used to group multiple events in the network. The framework used for defining time-based events in network data is scalable to various data sources. Recently, Loet Leydesdorff et al12 have developed a new layout technique based on multidimensional scaling for dynamic network layouts. Unlike previous approaches, such as SoNIA, they do not use linear interpolation between independent static frames. In the new layout algorithm proposed, the authors specifically add a parameter to control the change and movement of nodes over time. This parameter in turn controls how much the layout changes between subsequent frames and can be tuned to retain the mental model of the layout between different frames. This method is implemented in Visone63 a publicly available software system for visualising and analysing social networks. Despite the interest and advances from the visualisation community in dynamic social networks, the number of available systems used by social network scientists that handle dynamic networks is still lagging behind the developments reported from visualisation. Perhaps, the main reason for this is that these advances are developed in prototype systems and not extended into comprehensive general purpose systems that are widely available. Crowdsourcing experiments As part of our experiment, we employ a service provided by Amazon called Mechanical Turk (Mturk).64 MTurk 5 Farrugia and Quigley allows requestors (the employers) to outsource, or better crowdsource, small, stand alone tasks to a pool of workers. The tasks are typically small enough to be completed in a few seconds or minutes, and each task is usually paid in US cents. In the site’s vernacular, the tasks are called HITs – Human Intelligence Tasks. On the basis of sample usage examples in the service’s documentation, the site was envisaged to help with tasks such as image categorisation, survey responses and website filtering. The site also provides a facility to filter the workers that can participate on tasks either by measures of past performance or general demographics such as the worker’s location. Users can also be requested to qualify for certain HITs by first completing a qualification task designed to test the suitability of the worker to complete those tasks. This provides a good way to filter different workers based on their skill set and knowledge. Studies on the demographics of Mechanical Turk workers65,66 report an evolving trend in the worker population. Currently the highest concentration of workers is around 56 per cent US based, about 30 per cent Indian and a minority from other countries. The gender has become more balanced at almost 48 per cent male, 52 per cent female, as opposed to a 60 per cent female population in the early years of the service. The average age of workers is 31.6 and the education level is high with 59 per cent of workers having a Bachelor degree or higher. The reasons survey respondents give for working on the site are predominantly for making money (especially among the Indian population), for fun and to pass time. Such a platform is attractive to researchers as it provides a set of participants that are easily available, willing to conduct small tasks such as those involved in experiments, and usually involves a lower payment and time overhead in the process of conducting the experiments. The platform has already been used with promising success in user studies67 and to validate the quality of machine translation68 where Mechanical Turk workers were asked to judge the quality of a translation. Recently, Heer and Bostock69 studied the suitability of MTurk as an experimentation platform for visual perception, and found that the platform has potential for conducting viable visualisation experiments at a lower cost, faster completion time and with more participants than typical laboratory studies. The seminal study on graphical perception by Cleveland and McGill70 was replicated on MTurk with identical results to the original study. In the replicated study on contrast,71 the authors find that web-based experiments might, in fact, be more representative of the general usage in day to day life, rather than results in a laboratory based on a single display. Among their findings, Heer and Bostock claim that long tasks are not suited to MTurk as they are more prone to ‘gaming’. On the basis of this insight, we split the whole experiment in smaller chunks, yet promote the completion of the entire set of experimental tasks by offering a bonus for a complete set of answers. 6 Experiment Setup and Materials In order to study the impact of the mode of presentation on understanding dynamic networks, we conducted two distinct yet related experiments. Before the actual experiments, the experiment design and material was tested with a pilot group of social scientists, as distinct from our subsequent computer science students and online test subjects. Following the observations from the pilot study, our first experiment is a group laboratory experiment with computer science students familiar with node link diagrams. In the second experiment, we expand the audience of the first experiment to unspecialised participants by conducting the same experiment on the MTurk platform as introduced in Section ‘Crowdsourcing experiments’. A within subjects factorial design was employed in both experiments. The independent variables tested were 2× visualisation Type (Animation versus Static Images), 2 × 2 Tasks (Global Network overview versus Local Individual node versus Specified time period versus Unspecified time period) 2× Density (Low density versus High density) and 2× questions. The total number of questions for each participant were 2 × 4 × 2 × 2 = 32. In addition to this, four questions were presented to each participant in the beginning of the experiment to help them familiarise themselves with the system and question types. This data was subsequently discarded from the analysis. Tasks When studying dynamic networks, scientists are interested in change that occurs at different time periods in the network. Change in the network can occur at two levels; at the global network overview level and the local individual node level. Overview level tasks, are tasks that require the analysis of the entire network, whereas individual actor tasks focus an individual actor or a small group of actors. On the basis of prior experience with analysing dynamic networks,13 an analyst typically starts by identifying the overall change in the network before drilling down to investigate the actors that are responsible for that change. This task distinction is also applicable to static social network analysis and is synonymous with overview and detailed tasks when analysing attribute data. The temporal search space of a dynamic network is another important variable when studying dynamic networks. In a similar manner in which a network can be studied from an overview perspective and a local node perspective, the temporal search space can either be global across the entire time period or else localised to focus on a specified time period. Entire time period tasks question the complete data collection period of the network whereas specified time period tasks, question the network (or individual actor) at a specific point in time, for example in weeks 2, 3 and 4. Each of the first two task variables can be combined with tasks from the second category to Effective temporal graph layout Table 1: Examples of experiment questions for each task Network overview – Specified time period (Connection) Network overview – Entire time period (Density) Actor detail – Specified time period (Degree) Actor detail – Unspecified time period (Connection) Find a node that has only one connection on weeks 3, 4 and 5? Which week has the least number of connections? How do the number of connections of Node 1 change in weeks 2, 3 and 4? For how many weeks does the connection between Node 1 and Node 6 last? obtain network level tasks within a specified time period, network tasks throughout the entire time period, individual tasks with a specified time period and individual tasks throughout the entire time period. To formulate the user questions, we selected a number of tasks related to graph visualisation72 and formulated them in a dynamic context. For the individual actor tasks we developed questions on actor degree, path length between actors and general questions about the creation of new links, and the removal of old links. As a purely dynamic task for individual actors, we employed tasks on transitive relations between actors. For the overall network tasks we asked questions about network density and the number of groups observed throughout the entire network. An important network wide observation in dynamic networks characterises how actors form new ties, or change their linking preferences over time.15 As a pure dynamic concept question in this case, we included questions on homophilious patterns of connection between actors in the network. The principle of homophily states that new ties are more likely to form with other actors that resemble oneself.73 An example of a question on homophily is ‘Throughout the 6-week period, green nodes tend to form more connections with blue nodes, than with other green nodes?’. Other examples of questions for each task type are described in Table 1. The pilot questions asked to the social science students and subsequently the computer science students were questions that assumed a prior knowledge in interpreting social network or graph drawings. As MTurk workers might not necessarily be familiar with node link diagrams and graph concepts, questions about paths were removed and reformulated into other forms of questions. Also, based on the results of the pilot study and the first laboratory user study, questions on transitivity proved to be quite challenging, therefore they were removed from the set of question types for the MTurk. The four different types of question types used for MTurk included questions about connections, density, degree and homophily. The questions did not use any technical terms and were all phrased in terms of the connection patterns of the nodes. Hypotheses In this work, we ground our hypotheses on previous work on animation and the effect of motion on perception and apply it to social network visualisation. Hypothesis 1: Network overview tasks and no specified time period– Based on the work on the benefits of motion in overall tasks by McGrath and Blythe,47 we hypothesise that animation will be beneficial, therefore (a) faster and (b) more accurate, in analysing network overview task without a constraint on the time period. Hypothesis 2:Individual Actors and no specified time period – The positive results of animation used for transitions57 suggests that animation can be useful to facilitate the retention of context between different visualisations. If this notion is extrapolated a step further, one can hypothesise that animation may be beneficial, therefore (a) faster and (b) more accurate, for tasks that require the user to follow a node throughout the network. Hypothesis 3: Specific time period tasks – For localised tasks constrained by a time period, the overhead of interaction with the video for searching, stopping and pausing might result in animated presentations constrained by time being (a) slower than static versions. The interaction overhead however should not impact the (b) accuracy of the results. Hypothesis 4: Network density – Following Robertson et al,55 we hypothesise that the average time to respond will be (a) faster and (b) more accurate in lower density networks than in higher density networks. We are unsure about the effect of density and representation on each other, to this effect, we would like to investigate if the performance in the different representations is effected by the density of the network. Data sets Two data sets with different densities were used in this study. The first network (shown in Figure 1) is a low density, manually created network containing nine nodes across six different time periods. In this network, each node has one attribute encoded by shape and colour. The second network (shown in Figure 2) has a higher density and consists of 32 nodes, 2 node attributes, 2 types of edges and 6 time periods. In the visualisations, one of the node attributes is represented by colour and the other by shape. The second network is adapted from van Duijn et al74 who studied the evolution of friendship among a class of freshmen. We extracted part of this network that included only relationships of strength 1, 2 and 5 and removed edge direction. The reason for this simplification is to make the network more visually legible by removing edges of low semantic significance. Throughout, the experiment no reference was made to the source of the data in order to avoid any bias by participants who might have been familiar with the data set. 7 Farrugia and Quigley coordinates and each subsequent frame uses the node coordinates of the previous frame as a starting point. This technique was used as it has been found to be the most effective to generate these type of networks in previous studies.17 The layouts generated by the algorithm were improved by manually positioning some of the nodes to increase readability and retain context between time frames. SoNIA34 was used to generate all the visualisations and movies. Web-based experiment system Figure 1: Low density network week 4. Figure 2: High density network week 4. Network images and animation The two data sets in the study are presented to the participants as static network images and as animated movies. The two representations have an identical graph layout, colours and labels. In fact, the static images are taken from the six keyframes in the movie, one at each time point in the network. The movies were created by interpolating the positions of the nodes between the six keyframes used as static images. The movies for each network were generated at three speeds, normal (182 frames), slow (242 frames) and slower (302 frames), according to the feedback received during the pilot study. The algorithm used for generating the layouts is the Kamada-Kawai algorithm.75 To minimise the displacement of node position, chain-based anchoring is used between the successive layouts. In this approach, the first frame of each network is generated using random node 8 An online web-based system was specifically designed to run the experiment. The system is compatible with all major browsers with Flash and Javascript enabled to watch the movies and collect time measurements, respectively. The visual representations were displayed on the left, while the question was displayed on the same page to the right. The static images were laid out in a 3 × 2 grid, with each of the static images having the same size as the video (see Figure 3). Participants answered questions either by inputting a number in a text box, or using radio buttons for Likert scale questions. An option to skip a question was available if the participant was unable to answer. For any skipped questions participants were asked to supply a reason for skipping the question. The questions of the experiment were ordered using a Latin square design. The interaction techniques of the video animation were limited to play, pause and search for specific time position on the time bar. Pausing the video and clicking on the timeline enabled participants to move directly to a specific point in time. Participants were made aware of this functionality but no effort was made to promote this mode of interaction. Participants could also select their speed preference for the movie playback by clicking on the speed mode under the movie (see Figure 4). At the end of the experiment, each participant was presented with a list of adjectives and asked to select the top five words applicable to each representation. The list contained 100 adjectives and was equally split between positive adjectives and negative adjectives. The lists of words were randomised for each participant but kept the same for each representation. This qualitative data was gathered in an effort to understand the sentiments and impressions the participants experienced while conducting the experiment. Although a representation mode might be more effective in terms of measurable quantities such as accuracy and timeliness, there are other less easily quantifiable aspects that might provide a reason for using alternative presentation modes. Furthermore, a checklist of terms tends to encourage feedback more than simple open-ended questions on user preferences and this was reflected in the number of answers received for the adjective list as opposed to the open-ended question on the reason for representation preference. The list of adjectives also attempts to understand more subtle reasons for a preference towards a Effective temporal graph layout Figure 3: Web-based experiment system showing static image questions. representation by prompting the users with a list of emotions that they might not have necessarily thought about expressing but know that they have experienced. Adaptation of the lab System to MTurk The original system that was developed for the laboratory group study was adapted for MTurk. All the primary materials used in this experiment, images, videos, data set and experimentation system were kept the same. A minor addition was made to the experiment system to allow the users to pause the experiment. In a laboratory setting, the surrounding environment is designed to minimise external distractions and users are visually monitored. When an experiment is being run over the internet in a remote location, the experimenter loses all control of the surrounding environment and distractions. To minimise the effect of distractions on the timing of the results in performing the experiment, a pause button was added to the experimental system to allow the user to pause the timer. When the pause button is pressed the experiment materials are not displayed until the user presses the resume button to continue the experiment. The experiment still timed out after 30 min of inactivity, therefore if someone abandoned the experiment for a long time the data was discarded. Participants In the experiments different sets of participants were used. In the first experiment, the participants were seven males (average age: 29, min: 23) postgraduate research students in computer science, all of whom were familiar with node link diagrams. The student background was known and all the participants were previously screened before the experiment in terms of suitability for the tasks. In the second experiment, the participants were anonymous workers on Mechanical Turk. From the collected gender and age information, there were 26 females (average age: 29, min: 19, max: 57) and 33 males (average age: 30, min: 17, max: 58). In this case, there was no control over the participants and participants did not necessarily have knowledge of node link diagrams, social network or computer graphs. As reported in other studies a large proportion of MTurk workers have a high level of eduction,66 therefore understanding the basic principles behind node link diagrams was probably not difficult. The HITs submitted on MTurk were entered in four batches of 24 questions each, over a period of 11 days. Out of the 96 requests issued, 12 were rejected outright and reissued for a total of 108 submissions. From the 108 submissions, 48 submissions were discarded because they were completed by the same person more than once. These multiple submissions, identified by the unique worker id, were completed by seven workers. In this 9 Farrugia and Quigley Figure 4: Web-based experiment system showing animated image questions. case, only the first submission was taken into consideration and the others were discarded. From the number of submissions only three people did not complete the experiment in full. In order to have standardised results and retain the question ordering and within subjects design, these three submissions were discarded. All the submissions were screened for obvious signs of gaming of the system such as answering quickly and inaccurately. After cleaning the data, the total number of participants for the second experiment were 57. Laboratory group experimental method The experiments for the group of computer scientists were held in a classroom environment with each of the seven participants conducting the experiment simultaneously. Before each experiment, a brief tutorial was presented to explain the use of the system. A set of four test questions, each with a different representation and 10 task type were given to make the participants familiar with the system. Participants were encouraged to experiment with the system controls during these four example questions. MTurk experimental method The actual experiment platform was hosted on our servers to ensure maximum control on the presentation of content, collecting accurate timing data per question and control the question order to conform to the original Latin Square design. The MTurk HIT consisted of a description of the task with a link pointing to the experiment system and a passcode to uniquely identify each HIT. Upon finishing the experiment each participant was given a code to input in MTurk to finish the HIT. In order to keep in spirit with the small discernible tasks forming HITs on MTurk, we provided the option for the participants to finish the experiment early after Effective temporal graph layout completing each batch of four tasks. Each batch of four questions was worth 0.10c, and there was a total of nine batches. After each batch of four tasks was completed by the participant, a screen showing the monetary reward up to that point and asking if the user wants to continue was displayed. In order to incentivise the participant to complete the entire experiment, the participant was notified of a US $1 bonus if the entire experiment was completed. If the participant opted out of the experiment a question asking the reason for opting out was presented followed by the final qualitative questions. After the participant finished the experiment (or decided to stop prematurely), a passcode reflecting the level of completion was provided. We did this to prevent publishing the entire experiment as one long task with a high reward possibly attracting workers that want to ‘game’ the system and gain a ‘large’ some of money for a little reward.69 The MTurk platform does not provide a way to bundle a group of HITs together as a HIT can only be either rejected or approved. For this reason, all the HITs were published with the minimum amount of money for each HIT (0.10c) and not with the total amount, as otherwise people who do not complete the entire experiment will still get the maximum amount of money. The monetary balance based on the completion of the experiment was transferred to the participant in the form of a bonus in MTurk. When conducting an experiment remotely the experimenter has no control over the selection of participants, the surrounding environment and the computer setup of the participants. In order to ensure that the participants had a computer setup that conforms to the minimum requirements of the experiment, each remote participant had to pass a screening task before being able to participate. The screening task text provided the basic information necessary to interpret node link diagrams and answer the experiment questions. Each participant then had to pass a simple test that tested both the basic understanding of node link diagrams and also ensured that each participant had the minimum requirements to conduct the experiment (that is, watch a movie using Flash). because of the small number of participants, these results are not statistically significant. In the second experiment the same trends can be observed, but in this case most of the results are statistically significant. The performance of both groups of different users suggests that throughout most of the experiments, questions answered using static presentation were faster and sometimes more accurate than questions answered using animation. In general, this difference is stronger in measures related to time than measures of accuracy. In the laboratory experiment, questions with a static mode of presentation were answered faster (M = 56.45 s) than those with an animated presentation (M = 62.22 s). This difference however is not significant. In the second experiment, the t-test did reveal a significant difference in time to complete tasks between static and animated displays, t(904) = 9.154, (P < 0.001). This indicates that the average time to complete tasks using static representations (M = 35.16 s) was significantly lower than the average time to complete tasks using animation (M=47.55 s). The average time to complete is lower in the second experiment because the questions were simplified and in general were easier to answer. The experiment was designed this way due to the remote experimentation platform use and the anonymous possibly non specialised participants, taking part in the second experiment. The total accuracy in the first experiment was 85 per cent, whereas the accuracy of the second experiment was 86 per cent. The accuracy is sufficiently high in both cases to indicate that the participants were generally able to answer correctly and the questions were not too difficult. Participants in the second experiment who might have had no previous experience with node link diagrams were still able to answer most of the questions correctly. The similar accuracy rating of the participants suggests that the simplification of the questions for the second experiment using a non-specialised audience was appropriate and adequate. The number of errors made when animated presentation was used were higher (59 per cent of the total in the first experiment and 56 per cent of the total in the second), and the difference was significant t(904) = 2.660, (P < 0.01). Results In this section, we report on error rates, time to complete task, and descriptive preferences of the participants. The paired samples t-test was used to analyse both measures. General observations In the two experiments with the two different user groups the main results obtained are generally consistent. In terms of results, the first laboratory experiment with the smaller computer science students user group acted as a prequel to the larger experiment conducted with an anonymous audience. Most of the results of the first experiment are indicative of the general trend, however Task differences The 2 × 2 tasks for each group were cross tabulated and analysed in terms of average number of errors and average completion time. Once again both experiments reported similar results, in that static images were more effective in both time and accuracy. Figures 5 and 6 give an overview of the results in graphical format and Tables 2–5 give the results in tabular format. Network overview tasks with no specified time period Throughout the conditions tested in both experiments, the only measure that reported a positive difference for 11 Farrugia and Quigley Figure 5: Experiment 1 result overview. Figure 6: Experiment 2 result overview. Table 2: Average time to complete task – experiment 1 Table 4: Average time to complete task – experiment 2 Average time to complete Time Representation Animated Static Average time to complete Time Representation Animated Static Actor detail No specified time Specified time 41.10 56.28 36.31 53.15 Actor detail No specified time Specified time 62.83* 37.99* 44.81* 32.37* Network overview No specified time Specified time 71.01 77.42 70.66 67.66 Network overview No specified time Specified time 48.14* 42.88* 35.39* 29.65* ∗ Significant differences between averages. * Significant differences between averages. Table 5: Average error per task – experiment 2 Table 3: Average error per task – experiment 1 Average no of errors Time Representation Animated Static Actor detail No specified time Specified time 0.04 0.13 0.03 0.04 Network overview No specified time Specified time 0.21 0.31 0.28 0.13 Average no of errors Time Representation Animated Static Actor detail No specified time Specified time 0.25 0.14 0.22 0.11 Network overview No Specified Time Specified time 0.16* 0.06 0.11* 0.03 ∗ Significant differences between averages. animation was the accuracy measure for network overview tasks with no specified time period (Hypothesis 1b) in the first experiment. In the first experiment, animated images M = 0.21 reported a lower error rate than static images 12 M = 0.28, yet this difference is not significant. For the time to respond, the means are almost equal (M = 71.01 s) for animated and (M = 70.66 s) for static images and the difference is also not statistically significant. Effective temporal graph layout The only time the results between the experiments show different trends is in this task. Network wide questions without a specified time period were answered faster using static images (M = 35.39) than using animated movies (M = 48.14), t(226) = 3.815, P < 0.001. Here, questions answered using animated movies had a higher error rate M=0.16 than those answered with static images M=0.11. In this case, the difference is also significant t(226) = 1.95, P < 0.05. The results obtained from the second experiment contradict our first Hypothesis (1a and 1b), where animation is believed to be faster and more accurate in tasks on network overviews and no specified time. statistically significant. In the second experiment, the average time to reply in the dense networks is M = 47.57 s whereas in the lesser dense network it is M = 35.70 s, t(903) = −8.559 , P < 0.001. In case of errors, the difference is 0.13 in the lesser dense network and 0.14 in the dense network. We also conduct an RM-ANOVA analysis on the time measure of density data of the second experiment, to study the relationship between density and representation. From this analysis, we observe an effect of both density F(1448) = 70.48, P < 0.001 and representation F(1448) = 64.49, P < 0.001, but there was no significant interaction between density and representation. Individual actors and no specified time period In the first experiment, the results show that static images were both faster (M = 36.31 s for static, M = 41.10 s for animated) and had a smaller error rate (0.03 for static, 0.04 for animated) than animated movies, however both differences are not statistically significant. In the second experiment, a significant difference confirming that static images (M = 44.81s) are faster than animated movies (M = 62.83) is reported t(226) = 6.017, P < 0.001. The difference in terms of error rate is not statistically significant P < 0.286. This result also contradicts our initial hypothesis related to speed (2a), but is inconclusive on the second part of the hypothesis related to accuracy (2b). For tasks querying individual actors where the time condition is specified, the difference is significant for the timing measurement t(226) = 3.493, P < 0.01) but not for the error measurement. Participants’ comments and preferences Specific time period tasks In the first experiment, both the response time (M=60.41s for static, M = 66.85s for animated) and the error rate of static images (M = 0.08 for static, M = 0.21 for animated) have better performance than the animated representation, for tasks with a specific time period specified. Again, the results in this experiment are not significant. In the second experiment, the result mirror the results in the first experiment but have different significance values. The tests revealed a lower time to respond in the case of static visualisation (M = 30.19 s) when compared to animated visualisations (M = 40.41 s), t(451) = 6.264, P < 0.001. For the error rate measurement, the differences between static (M = 0.07) and animated (M = 0.10) representations is statistically significant t(451) = 1.636, P < 0.103. This result confirms our third Hypothesis 3a and 3b. Network density In Hypothesis 4, we postulate that denser networks lead to more errors and take a longer time to complete. This hypothesis is supported in the second experiment for the time to respond but not for the number of errors. In the first experiment, both differences are not At the end of the exercise, participants were asked to comment on both visual representations and indicate their preference between the static and animated displays. From the participants who answered that question, six preferred animations and 22 preferred static images. Perhaps, the most informative comment on this can be expressed by one of the participants who wrote ‘I really liked the videos but they were difficult to use if you are being timed, I felt a little rushed with them and turned them to the slowest speed’. One other participant remarked that he preferred ‘Static mostly, for accuracy. But the video was useful for following a single node around more easily’. This was in line with our original intuition, yet the quantitative results did not confirm it. The main reasons in favour of static displays included; ‘easier to access information selectively’, ‘easier comparisons over time periods’, ‘easy to verify results’. Some participants reported that they had to stop the animation and manually browse through the static frames to register the change. This behaviour was also noticed during the running of the laboratory experiment. In the words of one participant, ‘when trying to compare the transition between two weeks, it was very useful to be able to flip back and forth between two frames in the timeline and have them presented one on top of the other as in the animation.’ Some MTurk workers also commented positively on the nature of the HIT and expressed encouragement towards publishing other similar HITs. One participant also remarked on the novel structure of the HIT; ‘Hit was fairly enjoyable to complete. I like how it was broken up into four parts each, and I was able to stop and collect the money I earned without having to do all the questions if I didn’t want to. More hits broken up like that would be great!’. Some workers were even interested in the experiment and asked questions on the purpose of the experiment and the outcome (‘I wonder what you were really testing. It was interesting stuff anyway.’). It would probably be a good idea for future experiments to provide more information on the publications resulting from the experiment to create a stronger bond with the participants. 13 Farrugia and Quigley site words to the ones used for describing animation can be observed. The visualisation also clearly illustrates that participants preferred static images for their efficiency. Discussion Figure 7: Word cloud for adjectives selected for static representations. Figure 8: Word cloud for adjectives selected for static representations. In order to get a general impression of the sentiments that the visualisations elicited from the participants, each participant was asked to choose the top five words from a list of 100 words equally divided between positive or negative sentiments. Figures 7 and 8 show the list of static and animated words chosen by the participants respectively. In the figures, the size of the words corresponds with the number of times the word was chosen, whereas the locations of the words are random. The sentiments expressed for the respective representations are evident from the two diagrams. Most participants felt that animated images were harder and more difficult to use than the static alternatives. Participants also remarked on the time consuming nature of animation, a measure which was clearly reflected in the quantitative results. Time is of essence to workers on Mechanical Turk, since the longer a task takes the less money per HIT they can make, therefore this sentiment is well understood. Conversely, when one looks at the sentiments expressed for static images, oppo- 14 Similarly to other experimenters who were the first to test animation in a field that considered animation to be the natural way of representing temporal information,50 we were surprised at the overwhelming evidence in favour of static images. From the results, it is clear that for these experiments the static representations allowed participants to perform the tasks faster, and in some cases more accurately than animated representations. While for some tasks this was expected from the original hypotheses, most of the hypotheses on the benefits of animation were contradicted. In this discussion, we shall attempt to outline and understand some possible reasons for this. One possible reason is that interaction with animated displays was very limited, therefore the full potential of using animation was not exploited. Previous studies particularly in algorithm animations52 and our previous experiences43 have observed that interaction with animated displays is crucial during analysis. In the experiments, animation interaction was limited to pause, seek and play operations on the movie, rather than the interactions typically available in a complete interactive system. Time constraints and a desire to limit confounding factors in our experiments limited the use of such a complex system in the experiments. When animating one is effectively interpolating between many different frames creating many in-between or artificial graph layouts which in reality do not exist. If a user is interacting with these animations by starting and stopping the animations the search space is much larger because of the added interpolated frames. In this case, when a user is searching a video it takes much longer than searching a static image that quickly provides access to the final solution of the graph. A possible suggestion in this case, considering that some participants admitted that they used the animation like a static image by browsing the video, would be to let the user flip between keyframes and only animate when the play button is on. This means adding an additional interaction parameter to allow the user to browse between timeframes, slideshow style. Alternatively, keyframe locations can be highlighted on the video time line. Another possible reason is that most of the tasks in the experiment were low level topological tasks that require careful analysis of the network instead of a general overall perspective. In our first hypothesis, we believed that animation will be beneficial for general overview tasks, however this hypothesis was contradicted in both time and accuracy measures. While the network tasks were real overview tasks the questions might have been too specific, requiring the participants to look carefully for Effective temporal graph layout answers instead of trying to identify a general trend. General overview questions might also benefit from a layout scheme that emphasises the trend with respect to the position of the nodes. One of the participants commented that questions answered using animation felt like trying to hit a moving target. This comment might give some insight why our second hypothesis, where we thought that animation will be beneficial in tracking a single node, was contradicted. In our animation all the nodes were moving throughout the whole animation, thus possibly distracting the attention of the participants. One needs to be very careful when using animation, and it is probably better if animation is used sparingly rather than throughout a whole visualisation. Selectively animating parts of the network to exploit the pre-attentive nature of motion can prove to be more effective then simply animating the whole network. Interestingly, in our tests on density, the number of errors in the denser network was not significantly higher than the number of errors in the lower density network. One aspect of density that was not tested in this experiment was time density. Both networks spanned a period of six weeks, which resulted in 6 individual images and six key frames of video. It will be interesting to test animation on a higher time density network, say for example spanning a period of 20–200 time points. In such a case, looking at 20–200 different images of the network might be impractical as static images consume plenty of screen space, whereas animation or a flipbook may be easier to control as they can be contained in one screen. One problem with this sort of experiment would be the time required to run the experiment and the necessary precaution not to cognitively overload participants with a task that is too difficult. In this respect, we believe that a more ethnological type of study is probably a more realistic approach to study the analyst’s toolbox. The quantitative results from the measurements support the comments written by the participants in the final survey. This result is noteworthy because it informally validates the positive benefits of static images. The major complaint for animation, as can be clearly seen from the graphic in Figure 8, is that animation is hard to use. In our laboratory-based studies, some of the participants felt that animation took more time to become familiar with. From our word analysis, we found that participants in the laboratory tended to be more careful in selecting negative sentiments than anonymous participants, especially with respect to animation. Although for most tasks static representations were generally more accurate, the results were for the most part not statistically significant. We believe that there are still possibilities for using animation as an analytical tool, especially if animation is used selectively and with care. For example, animation can be useful to illustrate dynamics when presenting solutions, aided by additional voice over explanations of the network dynamics. Participants used positive sentiments such as ‘exciting’, ‘engaging’ and ‘fun’ when describing animated representations. Using remote web-based services for experimentation Web-based experimentation and web-based workers give the researcher the opportunity to reach a diverse participant group that extends beyond the typical sample of students typically available within a university context. This setup is a good stepping stone to start generalising certain types of experiments in a possibly more authentic setup than a laboratory space. Although we are still at a very early stage of understanding the benefits, drawbacks and possibilities that these web-based platforms can provide for research, we are encouraged by the results obtained and the quality of work obtained from the participants using the site. We caution other researchers to structure their experiments carefully and to consider any novelty effect in future uses of such a platform in experimental design. From our experience, we found that people who started the experiment were willing to complete the whole experiment, even if they were given the option to opt out and claim money at regular intervals. It will be interesting in future studies to measure the effect of the size of the bonus on the completion rate of the experiments. In some cases, the same participants completed the experiment several times in a row, giving the impression that workers on the opposite side are working in a production line. Workers on the site are accustomed to the repetitive nature of the tasks typically found on the site, which makes running longer experiments practical. In hindsight, the number of tasks run in the experiment could have been extended for the web-based experiment. One area that we were sceptical about during the design of the experiment with Mechanical Turk was in the collection of timing data from a remote experiment without the ability to control distractions in the environment. In reality, the timing data collected was well within the expected range of task time completion and there were no obvious instances where the time kept on running for a long time that was disproportionate to the task length. We can conjecture that the fact that the participants were made aware of the importance of timing and the addition of the pause button helped to ensure more accurate records. Apart from a few users who completed the experiment multiple times we did not find any instances of trying to game the system by answering quickly and inaccurately. We were careful to take into consideration the recommendations by other researchers who detected instances of gaming, such as keeping the initial reward relatively low not to attract undesirable participants and pre-screening the workers. It could also be the case that workers are learning that if they try to abuse the system in this way their employees can easily detect their ploy and reject the tasks accordingly. 15 Farrugia and Quigley We believe that there is promise in web-based experiments and that this experiment platform and traditional laboratory experiments can complement each other. The obvious benefits of participant resourcing, low cost and relatively fast data collection times make remote web systems attractive for the researcher. Naturally, the results obtained from this platform need to be interpreted in the context of the platform and care must be exercised in ensuring that the data is clean before being used for analysis. Future Work and Conclusions When interpreting networks it is impossible to be completely independent of the graph layout used to generate the images. As previous studies have shown23,37 the graph layouts have a significant effect on the interpretation and understanding of the graph structure. For this experiment, we use a layout algorithm that was originally designed for static networks, yet there are still very few graph layout algorithms designed specifically for dynamic networks, that are available in social network analysis toolkits. Furthermore, in these experiments the dynamic network visualisation was restricted to node link diagrams. There is further scope to look at different visual representations for dynamic network data that extends beyond the current paradigms of network visualisation, perhaps even being designed with an eye towards exploiting the power of motion in subtler ways. As we have discussed, there is scope for extending this study to test other criteria such as time density and perhaps investigate the difference between structural and analytical question formulation for tasks. Furthermore, more interactive features can be added to the animated displays to study the effect of interaction on network animation. Improvement in dynamic network layouts will undoubtedly be of great help to dynamic network visualisation irrespective of the representation medium used to visualise the network. The aim of this user study was to understand the strength of static images and animated movies to analyse dynamic social networks. We found that static images provide better performance than their animated counterparts. The results obtained though do not justify giving up further studies on animated representations. Nevertheless, improvements and further studies are required to understand better the beneficial uses of animation in applied social network analysis. References 1 Doreian, P. and Stokman, F. (1997) Evolution of Social Networks. Amsterdam: Gordon and Breach Publishers. 2 Christakis, N. and Fowler, J. (2008) The collective dynamics of smoking in a large social network. New England Journal of Medicine 358(21): 2249. 16 3 van Duijn, M., Zeggelink, E., Huisman, M., Stokman, F. and Wasseur, F. (2003) Evolution of sociology freshmen into a friendship network. Journal of Mathematical Sociology 27(2): 153–191. 4 Xu, J., Marshall, B., Kaza, S. and Chen, H. (2004) Analyzing and visualizing criminal network dynamics: A case study. Proceedings of the Second Symposium on Intelligence and Security Informatics Springer. 5 Moreno, J.L. (1941) Foundations of sociometry: An introduction. Sociometry 4(1): 15–35. 6 Freeman, L.C. (2000) Visualizing social networks. Journal of Social Structure 1. 7 Eades, P., Marks, J., Mutzel, P. and North, S. (1997) Graph-drawing contest report. In: G. Dibattista (ed.), Graph Drawing. Berlin, Heidelberg: Springer, pp. 423–435. 8 Fekete, J., Grinstein, G. and Plaisant, C. (2004) InfoVis 2004 Contest: The History of InfoVis. IEEE, http://www.cs.umd.edu/ heil/iv04contest/. 9 Grinstein, G., Plaisant, C., Laskowski, S., O’ Connell, T., Scholtz, T., and Whiting, M. (2008) Vast 2008 Challenge: Introducing Mini-challenges. Visual Analytics Science and Technology, 2008. VAST ’08. IEEE Symposium. 10 Erten, C., Harding, P., Kobourov, S., Wampler, K. and Yee, G. (2003) GraphAEL: Graph animations with evolving layouts. Lecture Notes in Computer Science 2912: 98–110. 11 Bender-deMoll, S. and McFarland, D. (2006) The art and science of dynamic network visualization. Journal of Social Structure 7(2). 12 Leydesdorff, L., Schank, T., Scharnhorst, A. and De Nooy, W. (2008) Animating the development of Social Networks over time using a dynamic extension of multidimensional scaling. El Profesional de Informacion 17(6): 611–626. 13 Farrugia, M. and Quigley, A. (2009) TGD: Visual data exploration of Temporal Graph Data. In: K. Borner and J. Park (eds.) Visualization and Data Analysis 2009, Vol. 7243, No.1, San Jose, CA: SPIE, p.11. 14 Archambault, D. (2009) Structural differences between two graphs through hierarchies. Proceedings of Graphics Interface 2009. 15 Powell, W., White, D., Koput, K. and Owen-Smith, J. (2005) Network dynamics and field evolution: The growth of interorganizational collaboration in the life sciences. American Journal of Sociology 110(4): 1132–1205. 16 Pajek. http://pajek.imfm.si/doku.php. 17 Moody, J., McFarland, D. and Bender-deMoll, S. (2005) Dynamic network visualization 1. American Journal of Sociology 110(4): 1206–1241. 18 Wasserman, S. and Faust, K. (1994) Social Network Analysis. Cambridge UK. 19 Herman, I., Melançon, G. and Marshall, M. (2000) Graph visualization and navigation in information visualization: A survey. IEEE Transactions on Visualization and Computer Graphics 6(1): 24–43. 20 Di Battista, G., Eades, P., Tamassia, R. and Tollis, I. (1999) Graph Drawing; Algorithms for the Visualization of Graphs. Upper Saddle River, NJ: Prentice Hall. 21 Kaufmann, M. and Wagner, D. (2001) Drawing Graphs: Methods and Models. Springer-Verlag: London, UK. 22 Di Battista, G., Eades, P., Tamassia, R. and Tollis, I. (1994) Algorithms for drawing graphs: An annotated bibliography. Computational Geometry: Theory and Applications 4(5): 235–282. 23 Purchase, H. (1997) Which aesthetic has the greatest effect on human understanding? In: G. Di Battista (ed.) Graph Drawing. Lecture notes in Computer Science, Vol. 1353. Berlin; Heidelberg: Springer, pp. 248–261. 24 Ware, C., Purchase, H.C., Colpoys, L. and McGill, M. (2002) Cognitive measurements of graph aesthetics. Information Visualization 1(2): 103–110. 25 Misue, K., Eades, P., Lai, W. and Sugiyama, K. (1995) Layout adjustment and the mental map. Journal of Visual Languages and Computing 6(2): 183–210. 26 Purchase, H., Hoggan, E. and Görg, C. (2007). How important is the ‘mental map’? – An empirical investigation of a dynamic graph layout algorithm. Lecture Notes in Computer Science 4372: 184. Effective temporal graph layout 27 Brandes, U. (2001). Drawing on physical analogies. Lecture Notes in Computer Science 2025: 71–86. 28 Brandes, U. and Wagner, D. (1997) A bayesian paradigm for dynamic graph layout. Lecture Notes in Computer Science 1353: 236–247. 29 Frishman, Y. and Tal, A. (2008) Online dynamic graph drawing. IEEE Transactions on Visualization and Computer Graphics 14: 727–740. 30 Diehl, S., Görg, C. and Kerren, A. (2001) Preserving the Mental Map Using Foresighted Layout. Proceedings of the Joint EurographicsIEEE TCVG Symposium on Visualization VisSym ’01, Ascona, Switzerland, Wien, Austria: Springer Verlag, pp. 175–184. 31 Lee, Y., Lin, C. and Yen, H. (2006) Mental Map Preserving Graph Drawing Using Simulated Annealing. Proceedings of the 2006 Asia-Pacific Symposium on Information Visualisation-Volume 60, Australian Computer Society, p. 188. 32 Davidson, R. and Harel, D. (1996) Drawing graphs nicely using simulated annealing. ACM Transactions on Graphics 15(4): 301–331. 33 Dwyer, T. and Gallagher, D. (2004) Visualising changes in fund manager holdings in two and a half-dimensions. Information Visualization 3(4): 227–244. 34 Bender-deMoll, S. and McFarland, D. (2006) The art and science of dynamic network visualization. Journal of Social Structure 7(2) 35 Bridgeman, S. and Tamassia, R. (2002) A user study in similarity measures for graph drawing. Lecture Notes in Computer Science 6(3): 225–254. 36 Friedrich, C. and Eades, P. (2002) Graph drawing in motion. Journal of Graph Algorithms and Applications 6(3): 353–370. 37 Blythe, J., McGrath, C. and Krackhardt, D. (1996) The effect of graph layout on inference from social network data. Proceedings of the Symposium on graph Drawing GD ’95. London: Springer-Verlag, pp. 40–51. 38 Tversky, B., Morrison, J. and Betrancourt, M. (2002) Animation: can it facilitate?: International Journal of Human-Computer Studies 57(4): 247–262. 39 McFarland, D. (2001) Student resistance: How the formal and informal organization of classrooms facilitate everyday forms of student defiance1. American Journal of Sociology 107(3): 612–678. 40 Brandes, U., Kääb, V., Löh, A., Wagner, D. and Willhalm, T. (2000) Dynamic WWW Structures in 3D. Journal of Graph Algorithms and Applications 4(3): 183–191. 41 Ke, W., Börner, K. and Viswanath, L. (2004) Major information visualization authors, papers and topics in the ACM library. Proceedings of the IEEE Symposium on Information Visualization. Washington DC: IEEE Computer Society, p. 216.1. 42 Lee, B., Czerwinski, M., Robertson, G. and Bederson, B. (2004) Understanding Eight Years of Info Vis Conferences Using PaperLens. Washington DC: IEEE Computer Society, p. 216.3. 43 Farrugia, M. and Quigley, A. (2008) Cell phone mini challenge: Node-link animation award animating multivariate dynamic social networks. IEEE Symposium on Visual Analytics Science and Technology, 2008. VAST ’08; October, IEEE, pp. 215–216. 44 Correa, C.D., et al. (2008) Cell phone mini challenge award: intuitive social network graphs visual analytics of cell phone data using Mobivis and Ontovis. IEEE Symposium on Visual Analytics Science and Technology, 2008. VAST ’08; October. IEEE, pp. 211–212. 45 Perer, A. (2008) Using socialaction to uncover structure in social networks over time. IEEE Symposium on Visual Analytics Science and Technology, 2008. VAST ’08; October. IEEE, pp. 213–214. 46 Perer, A. and Shneiderman, B. (2006) Balancing systematic and flexible exploration of social networks. IEEE Transactions on Visualization and Computer Graphics, 12: 693–700. 47 McGrath, C. and Blythe, J. (2004) Do you see what i want you to see? the effects of motion and spatial layout on viewers’ perceptions of graph structure. Journal of Social Structure 5(2). 48 Ware, C. and Bobrow, R. (1985) Motion to support rapid interactive queries on node-link diagrams. ACM Transactions on Applied Perception (TAP) 1(1): 3–18. 49 Brown, M. and Sedgewick, R. (1985) Techniques for algorithm animation. IEEE Software 2(1): 28–39. 50 Stasko, J., Badre, A. and Lewis, C. (1993) Do Algorithm Animations Assist Learning?: An Empirical Study and Analysis. Proceedings of the INTERACT’93 and CHI’93 Conference on Human Factors in Computing Systems ACM CHI ’93; Amsterdam; The Netherlands; New York: ACM, pp. 61–66. 51 Kehoe, C., Stasko, J. and Taylor, A. (2001) Rethinking the evaluation of algorithm animations as learning aids: An observational study. International Journal of Human Computer Studies 54(2): 265. 52 Lawrence, A., Badre, A. and Stasko, J. (1994) Empirically Evaluating the Use of Animations to Teach Algorithms. Proceedings of the 1994 IEEE Symposium on Visual Languages, Citeseer, pp. 48–54. 53 Rosling, H. (2006) Debunking myths about the ‘third world’,. http://www . gapminder . org / videos / ted - talks / hans - rosling - ted2006-debunking-myths-about-the-third-world/. 54 Rosling, H. (2007) The seemingly impossible is possible, http://www.gapminder.org/videos/ted-talks/hans-rosling-ted-talk2007-seemingly-impossible-is-possible/ 55 Robertson, G., Fernandez, R., Fisher, D., Lee, B. and Stasko, J. (2008) Effectiveness of animation in trend visualization. IEEE Transactions on Visualization and Computer Graphics 14(6): 1325–1332. 56 Yee, K.-P., Fisher ,D., Dhamija, R. and Hearst, M. (2001) Animated Exploration of Dynamic Graphs with Radial Layout. Proceedings of the IEEE Symposium on Information Visualization. INFOVIS ’01. Washington DC: IEEE Computer society, p. 43. 57 Heer, J. and Robertson, G. (2007) Animated transitions in statistical data graphics. IEEE Transactions on Visualization and Computer Graphics 13(6): 1240–1247. 58 Elmqvist, N., Dragicevic, P. and Fekete, J.-D. (2008) Rolling the dice: Multidimensional visual exploration using scatterplot matrix navigation. IEEE Transactions on Visualization and Computer Graphics 14(6): 1148–1539. 59 Snijders, T., Steglich, C., Schweinberger, M. and Huisman,M. (2007) Manual for SIENA, version 3.1. 60 Ware, C. (2004) Information Visualization: Perception for Design. San Francisco, CA: Morgan Kaufmann. 61 UCINET. http://www.analytictech.com/ucinet/ucinet.htm. 62 Batagelj, V. and Mrvar, A. (1998) Pajek-program for large network analysis. Connections 21(2): 47–57. 63 Brandes, U. and Wagner, D. (2003) Visone: Analysis and visualization of social networks. Graph Drawing Software. Secaucus, NJ: Springer-Verlag New York, pp. 321–340. 64 Turk, M. http://www.mturk.com. 65 Ipeirotis, P. (2010) Demographics of Mechanical Turk, http://behindthe-enemy-lines.blogspot.com/2010/03/new-demographics-ofmechanical-turk.html. 66 Ross, J., Irani, L., Silberman, M., Zaldivar, A. and Tomlinson , B. (2010) Who are the Crowd-workers?: Shifting Demographics in Mechanical Turk. Proceedings of the 28th of the International Conference Extended Abstracts on Human Factors in Computing Systems ACM, CHI EA’10; Atlanta, GA. New York: ACM, pp. 2863–2872. 67 Kittur, A., Chi, E., and Suh, B. (2008) Crowdsourcing user studies with Mechanical Turk. Proceeding of the Twenty-Sixth Annual SIGCHI Conference on Human Factors in Computing Systems ACM, CHI ’08; Florence, Italy. New York: ACM, pp. 453–456. 68 Callison-Burch, C. (2009) Fast, cheap, and creative: Evaluating Translation quality using Amazon’s Mechanical Turk. Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing; Association for Computational Linguistics, pp. 286–295. 69 Heer, J. and Bostock, M. (2010) Crowdsourcing graphical perception: Using mechanical Turk to assess visualization design. Proceedings of the 28th of the International Conference Extended Abstracts on Human Factors in Computing Systems, CHI ’10; Atlanta, GA; New York: ACK, pp. 203–212. 70 Cleveland, W. and McGill, R. (1984) Graphical perception: Theory, experimentation, and application to the development of graphical methods. Journal of the American Statistical Association 79(387): 531–554. 17 Farrugia and Quigley 71 Stone, M., Bartram, L. and Consulting, S. (2009) Alpha, contrast and the perception of visual Metadata. Sixteenth Color Imaging Conference: Color Science and Engineering Systems, Technologies and Applications, Vol. 16, November 2008, Portland, OR: The Society for Imaging Science and Technology. 72 Plaisant, C., Lee, B., Parr, C., Fekete, J. and Henry, N. (2006) Task taxonomy for graph visualization. Beyond Time and Errors: Novel Evaluation Methods for Information Visualization (BELIV 06): 82–86. 18 73 McPherson, M., Smith-Lovin, L. and Cook, J. (2001) Birds of a feather: Homophily in social networks. Annual Review of Sociology 27(1): 415–444. 74 van Duijn, M., Zeggelink, E., Huisman, M., Stokman, F. and Wasseur, F. (2003) Evolution of sociology freshmen into a friendship network. Journal of Mathematical Sociology 27(2): 153–191. 75 Kamada, T. and Kawai, S. (1989) An algorithm for drawing general undirected graphs. Information Processing Letters 31: 7–15.