Visualisation of Linked Data – Reprise

1.Introduction

Linked Data is a key component of the Semantic Web vision. The ability for data consumers to adopt a follow your nose approach, traversing links defined within a dataset or across independently-curated datasets, is an essential feature of what we now call the Web of Data, enabling richer knowledge retrieval thanks to synthesis across multiple sources of, and views on, inter-related datasets. Since its early days, Linked Data (LD) promised to serve as a disruptor of traditional approaches to data management and use, promoting the push from the traditional Web of documents to this Web of data [3,4,40]. But the advantages of following LD principles have now become even more apparent [21,39], as the boundary between the data available on the (online) Web and users’ own (personal) data gets ever more fuzzy. Users in our increasingly data- and knowledge-driven world are becoming more and more dependent on applications that build upon the capability to transparently fetch heterogeneous yet implicitly connected data from multiple, independent sources. It has become clear that we must rethink how those users interact with the very large amounts of this complex, interlinked, multi-dimensional data, throughout its management cycle, from generation to capture, enrichment in use and reuse, and sharing beyond its original context.

The Linked Data community introduces its foundational concept as follows [30]:

At the simpler level of individual data consumers, lay (non-expert) users may eventually employ Linked Data as a portal to timely, context-based knowledge for everyday activities, doing so either implicitly or explicitly. Such activities could range from simple ones, such as comparing the price of groceries or planning a journey using public transport, to more complex ones involving important decisions, such as purchasing life insurance or a new home. LD has the potential to make such activities more efficient by giving users access to a rich Web of interlinked data relevant to the task at hand; however, this should not be accompanied by technical barriers that are difficult to overcome by a majority of users.

The design of user interfaces for LD [22,24,54], and more specifically interfaces that represent the data visually, play a central role in that respect. Well-designed visualisations harness the powerful capabilities of the human perceptual system, providing users with rich representations of the data. Combined with appropriate interaction techniques, they enable users to navigate through, and make sense of, large and complex datasets, help spot outliers and anomalies, recognise patterns, identify trends [7,49]. Visualisation can provide effective support for confirming hypotheses, and favours deriving new insight. Visual analytics systems complement human cognitive and perceptual abilities with automated data processing and mining techniques [9,39,47,55]. Such systems not only support the presentation of data, but processes for the full data management cycle as well, from data capture to analysis, enrichment and reuse.

However, visualisation in and of itself is not an absolute solution to the challenges inherent to complex data sense-making, analysis and reuse. To contribute to the realisation of LD’s potential, visualisations have to be tailored to specific tasks, and effectively support users in the performance of these tasks by providing visual representations of the data at a relevant level of detail and abstraction. Following Shneiderman’s mantra [49], users should be able to navigate from an overview of the data to detailed representations of specific items of interest. But the system should also provide relevant representations of the data at these various levels of detail [7]. This is a major challenge, considering the follow-your-nose approach to data exploration mentioned earlier, and given the potentially high level of heterogeneity of the interlinked datasets involved in most LD-backed applications. Only then will the results of visual exploration and analysis lead to new insights about LD, with an increased level of confidence in drawing conclusions from the data and the supporting visual analysis, and in decision-making based on the knowledge thus acquired.

We review the state of the art in Section 2, and look at increasing efforts to evaluate front-end tools with target end-users in Section 3. In Section 4 we discuss on-going research in Linked Data visualisation and the range of challenges tackled by contributors to the special issue, as part of the process followed in designing and building effective visual presentation and analysis solutions for Linked Data. We conclude in Section 5 with a vision for the role of visualisation in advancing research in, analysis of, and applications based on Linked Data.

2.The state of the art: Linked Data visualisation

Tackling the challenges associated with the representation of Linked Data and its visual exploratory analysis requires considering related, supporting technologies. Halpin and Bria [17], Klímek et al. [27], Mai et al. [37], Mitchell and Wilson [39], for instance, highlight reliance on data structuring, typically using ontologies as a backbone; the identification and encoding of links via ontology mapping and alignment; and the need to integrate such technologies into the visual discovery and analysis pipeline. Formal querying is a powerful tool for extracting specific information either from one or across multiple datasets. However, formulating any but very simple queries in SPARQL (the de facto RDF query language [52]), or any other formal query language, is tedious even for SW experts. It is also error-prone, especially with each additional graph and optional or required relationship [9,39]. Further, the textual results returned by query engines, which typically take the form of machine-readable RDF, are not easily interpreted by users, even if they have a good level of domain understanding. Even when considering expert users, meaningful overviews are required to obtain a good understanding of data structure, content and relationships for any but very small result sets. Predefined query templates serve as an aid here, and as a foundation on which more intuitive support for guided querying and question-answering may be built.

Valsecchi et al. [54] demonstrate the advantages gained in identifying suitable entry points to LD exploration, as a key initial step to improving access, especially for non-SW and non-domain experts. Benedetti et al. [2], Haag et al. [16] discuss the need to support visual querying to overcome technical challenges in the formulation of formal queries by lay users and, as complexity grows, even technical experts. Mai et al. [37] further demonstrate the added value in providing multiple visual perspectives that support the follow your nose paradigm when exploring query results across multiple, heterogeneous, albeit linked, datasets. Halpin and Bria [17] describe the process followed to map how Digital Social Innovation is being used throughout the European Union (EU) across multiple aspects of people’s lives, and measure the resulting social impact. They illustrate the benefits gained in employing SW technology to overcome several challenges – challenges posed by differences in structure, granularity and completeness of the data sources involved. Indeed, multiple sources have to be queried to capture all data required to create a DSI map (of the EU), a process that may be used to generate a linked open dataset synthesising this data.

Beyond tools that help initiate exploration by identifying entry points into LD, users should also be provided with support for identifying and selecting more focused resources. Depending on the use case, this might involve increasingly customised (or customisable) tools for in-depth exploration and analysis, along with support for reviewing content and sharing both this content and the results of discovery and analysis beyond its initial point of generation or use. Toward this aim, it may be instructive to follow best practice and verified frameworks and models for the design of visualisation and visual analytics such as in [47,49,55], adapting these where necessary for the specific case of LD and SW data in general. As in [9], we use Shneiderman’s seminal visual information-seeking mantra as a guiding principle in evaluating the usability and utility of visualisation-driven approaches and tools for different users and their tasks – looking at support for: (i) generating overviews on the underlying and related data; (ii) filtering out noise and less important data in order to focus on ROIs (Regions of Interest); and (iii) visualising ROIs in detail.

We review selected work in the SW community that illustrate advancement of the state of the art in Linked Data visualisation in the five years since the 2011 survey about LD visualisation [9], looking at models, approaches and tools that focus on one or more of the following visualisation-driven tasks:

Sections 2.1, 2.2 and 2.3 look at the roles of each of (i) ontologies, (ii) formal querying and exploration and (iii) content analysis and presentation respectively, and their role in the broader task of LD visualisation. However, the discussion that follows shows a good amount of overlap between all three aspects, reinforcing the benefits in integrating these component parts of the process and/or building seamless pipelines between them.

3.Evaluating usability & utility

Recognition of the need to support human end-users (in addition to the focus on machines) is receiving greater traction within the SW community, especially for applications targeted at lay users and even domain experts who are not necessarily technological experts. This is critical to extending reach of the technology beyond the SW community, and especially important as expectations of end users grow with advances in technology. Further, as data size and complexity increase exponentially with ubiquitous use of said technology [39,58, among others], the need for intuitive, user- and task-focused tools for navigating through the noise to identify and make effective use of relevant knowledge, within the user’s context of use and field of reference, increases. This is true for domain and technical experts and lay users, the latter of whom range from technophobic or low technology awareness, to those who routinely use a range of advanced technological devices and software in their personal, social and working lives.

One benefit in the pipeline or integrated approach to tackling requirements for LD visualisation is that the process itself serves as a knowledge retrieval task and a form of evaluation of both the approach and the tools and techniques that feed into it [see, e.g., 21,37,47]. The development team can only progress to the next stage in what is often an iterative pipeline or workflow if they have been successful in: (a) correctly identifying relevant resources; (b) extracting data and metadata that meet the user’s context and the requirements of their task or end goal; (c) preprocessing this data, where necessary, to feed into the visualisation tools available and the analysis required; (d) carrying out initial exploratory analysis, using analytical and/or mining models, statistical and/or visualisation-driven methods. These initial steps will confirm whether the input data and preprocessing will feed into solving the problem at hand, and therefore guide the identification of suitable approaches for more detailed analysis, or may need to be reviewed to improve the process or correct issues found.

We therefore now see an increase in usability evaluation of tools built on SW technology. Further, the open data initiative has spurred the sharing of what is typically considered personal or in-house research data (such as interim analysis and evaluation data) along with final analysis results, to encourage verification of research by third parties and contribute to benchmark datasets. Tietz et al. [53], for instance, make available online anonymised evaluation data for their tool refer. They evaluate usability of the text visualisation tool with a range of regular Web users, a quarter of whom described themselves as LD experts, while the others had passing to no knowledge of Linked Data. The study found that while the UI was, overall, seen as useful, and navigation sufficiently intuitive to allow both user types to discover the information sought, some UI features for revealing additional information were not recognised especially by the lay users. An interesting finding: criteria for categorising the data and proposing recommendations, a key component in obtaining more complete understanding of relationships revealed as users navigate through data, was assumed by some users to be the whole document, rather than the entity with the focus. This highlights a significant knowledge gap between SW experts and lay users, with the push in the SW from a Web of documents to a Web of data, and the finer granularity available with the latter.

Benedetti et al. [2] in what they describe as qualitative evaluation, involving both lay and technical users, report high satisfaction and accuracy across both user types for tasks requiring browsing of their visual schema summaries (91% with up to 5 min to complete), and successful task completion based on visual query generation (90% with up to 15 min to complete). It should be noted that the authors restrict the tasks to those that can rely on browsing the schema summaries, and do not compare their results with benchmarks, nor mention whether or not relevant benchmarks exist.

Valsecchi et al. [54] target technological experts and evaluate their tool only with users who have a Computer Science background. Participants worked with two maps generated from highly inter-connected DBpedia [12] data and very disconnected LinkedMDB [19,32] data, respectively. The evaluation solicited qualitative and quantitative information on intuitiveness and learnability of the UI, and how well it supported users in obtaining an understanding of data structure, including connectedness and content. Results were mixed: success and correctness in completing tasks ranged from 65% to 100%, with higher scores overall for the DBpedia layout. However, participants were able to discover only some of the key functions, including a subset of those required for navigation. A key finding was the importance of providing additional textual detail to complement and confirm the conclusions drawn from the visual overviews. This is not an unusual finding: Ware [56], for instance, illustrates how text may be used to complement visual thinking and understanding.

We conclude this section with a summary in Table 1 of prototypes built and new approaches described in this section and in Section 2, by comparing these according to a set of criteria for usable visualisation, derived from guidelines for interactive visualisation and visual information seeking found in [49], [see also 9]. It should be noted that the assessment is made using authors’ descriptions where an implemented prototype is not available publicly; such cases are highlighted with an asterisk after the tool name.

4.The special issue

The call for papers for a special issue of the Semantic Web Journal on “Visual Exploration and Analysis of Linked Data” was posted in June 2014. We received 13 formal expressions of interest and 11 complete submissions. Of these, 6 were accepted for publication, addressing between them the three key components identified in the call, that feed into effective Linked Data visualisation (see also Section 2).

Linked Data, open or closed, evokes discussions about big data, within the SW community and further afield, in government and in industry, due to its size, complexity, heterogeneity, dynamicity, distributed nature and differences in granularity, quality and veracity. But LD goes beyond simply being big, serving as an enabler with potential to address some of the key issues in managing vast amounts of data. While not required to follow a schema, a situation that contributes to challenges in its use, the inherent structure of LD supports identification of the links within a dataset and across to other independently generated data, enabling more seamless exploration and richer information discovery. The authors in this special issue highlight the importance of following guidelines and best practice for curating Linked Data as well as those for visual information presentation and analysis, to work toward realising the full potential of LD.

Fu et al. discuss the need to evaluate visualisation-based techniques used in ontology understanding, a task that contributes to the process of ontology mapping, which is in turn a contributor to linking independent datasets. The authors argue that such evaluation is key to the design of usable, intuitive, user-facing tools.

Nuzzolese et al. mine the link structure in LD to rank importance of data associated with a knowledge entity of interest, based on popularity with respect to the contributions of the crowd to encyclopaedic knowledge. The results are used to generate visual summaries about the topic of interest and support step-wise navigation to other related data. Bikakis et al. aggregate LD hierarchically based on its backing ontological structure, to improve navigation through the data using a variety of visualisation techniques built on top of the resulting aggregates. Scheider et al. harness the familiar metaphor provided by maps to improve navigation through LD, and hence, exploration and querying of LD with spatio-temporal attributes.

A number of one-off visualisation solutions exist. However, low to no reusability limits their usefulness. An important challenge being tackled is, therefore, the design of workflows, pipelines and integrated, multi-perspective solutions that tackle different stages in the complete Linked Data management cycle, typically focusing on data selection and pre-processing, initial exploration and a selection of tools for more detailed analysis of ROIs and for presenting the results of analyses. Brasoveanu et al. describe the task-centred process they followed to build a visualisation dashboard for decision support, comprising a set of coupled tools for aggregation and analysis of linked open statistical data. Del Rio et al. extend the five-star methodology for generating high quality LD. They define a seamlessness metric for visual analytics, with the goal of improving reuse of the interim and final results obtained during visual analysis, by embedding semantic data into the process and the analyses’ output.

Fig. 1.

Building queries using the click-and-select UI in SPEX, the visual query editor built to support non-technical end users exploring Linked Data.

Cartography is an oft-used metaphor when generating visualisations, as it takes advantage of users’ familiarity with and consequent ability to read maps to aid exploratory discovery. For example, a user may wish to explore Napoleon’s march to Russia in the War of 1812 [38]. The classic map requires mainly human perception to understand its content; expert knowledge in history and cartography may provide some advantages but are not strong requirements. However, making the visualisation and its content searchable in a digital data store requires additional annotations, both textual and using visual overlays, to be stored with the map. Scheider, Degbelo, Lemmens, van Elzakker, Zimmerhof, Kostic, Jones & Banhatti in Exploratory Querying of SPARQL Endpoints in Space and Time examine challenges of this type, faced by users in formal querying of data stores as part of the exploration and question-answering process. Even assuming such annotations exist, end users, whether technological experts or lay persons, require additional expertise to successfully retrieve information required to answer their questions; namely, the need for expertise in formal query languages such as SPARQL, at least basic understanding of data structure and content, and depending on the data types involved, understanding of complex data constructs.

The authors therefore look at two data types – spatial (physical or geographical space) and temporal (time) – that occur frequently across many domains. Scheider et al. employ space and time as features that may be used to explore individual data sets and as a bridge between data silos, by harnessing the everyday metaphor of maps coupled with a timeline. They also follow Shneiderman’s mantra [49] to define a set of design requirements to support intuitive, visual exploration. They aim to enable a broad range of users, that do not necessarily have much domain or technological expertise, to progressively build and edit queries as they explore linked datasets along attributes defining “space and time”.

Scheider et al. describe a qualitative usability evaluation carried out with six participants: two librarians, two PhD students in Geography, and two ITC lecturers. The goal of this study was to assess the extent to which their prototype SPEX, the SPatio-temporal Content Explorer (shown in Fig. 1), satisfies the design guidelines they specify. The authors report between 45 and 90 minutes for familiarisation with the tool using a printout of its help file, followed by 30 to 45 minutes to complete a set of exploratory tasks. Participants made good use of all query features, including functionality for reuse of previous results. However, none were successful at completing all tasks, with an overall success rate of 52%. The authors note, though, that even where participants were not successful in completing the tasks, in some cases they tried alternative strategies to answer the questions, with some of these alternatives leading to partial success or a correct answer(s).

Key challenges faced by the participants were related to the unfamiliar use of terminology and labeling of data – which, derived directly from the RDF, would be relatively easily interpreted by a SW expert, but did not translate easily otherwise, especially for the more ambiguous terms. A simple example is distinguishing between a map entity and the (visual) geographical map. Other challenges were due to user interface design issues such as what were found to be redundant steps to trigger events in the dialogs, and how to generate sub-queries using the timeline.

The authors conclude the paper with future directions for research to more effectively translate their design guidelines into more intuitive support for a non-technical, lay user audience.

Ontologies typically serve as a backbone for structuring Linked Data, and play a key role in defining links within a dataset and across datasets to other related data. Tools for constructing and exploring ontologies and especially for mapping and alignment of concepts and properties therefore contribute to effective management and (re)use of LD. Fu, Noy & Storey in Eye Tracking the User Experience – An Evaluation of Ontology Visualization Techniques recognise a gap between work to develop ontology visualisation tools and the usability evaluation required to ensure best practices are followed in the design and development of such tools. They highlight the resulting negative impact on the appropriation of useful and usable techniques that map to user expertise and knowledge requirements in ontology engineering and related knowledge management tasks, Linked Data being one obvious example.

Fu et al. use eye tracking to collect empirical evidence about usability in a comparative study of two of the most widely used visualisation techniques in ontology understanding: indented lists and trees. The authors argue that eye tracking, being able to directly capture user actions, provides more in-depth measures, at a micro level, to help determine why an approach or tool may be effective, as opposed to reliance only on traditional measures of performance during usability evaluation such as time to completion and success rate, that provide indirect means for capturing usability.

The study examined four hypotheses: (i) search, (ii) information processing, (iii) ability to reduce cognitive load, (iv) why a particular technique supports more efficient and/or effective use by particular types of users.

The study involved 36 participants, graduates and undergraduates from a wide range of disciplines within the Sciences. They were instructed to perform mapping tasks between two sets of ontologies; each session lasted approximately 2 hours. Participants were required to verify whether existing mappings were correct and complete (identifying missing mappings otherwise). Visualisations of each ontology were provided to aid task completion, which relied on exploration of each pair of ontologies, both created from standard benchmark datasets. One pair was simple and described conferences. The second pair from the biomedical domain, describing organisms, was more complex. The authors note deliberate effort to reduce participant bias due to familiarisation with the data domains and the visualisation techniques, in order to simulate as closely as possible challenges faced by novices (with no knowledge of SW technologies and especially ontologies). This restriction allowed measurement of the visualisation support for completing the task while reducing confounding factors, on the basis that domain- or SW experts would require less intuitive support to complete the tasks successfully. Further, the visualisations were generated to be as representative of current applications as possible, by implementing the most typical cues for navigation and browsing, focusing on ROIs, filtering and hiding data. Also, each task started with only the root node visible, so that participants had full control over their navigation paths.

The results, based on 31 participants (five were discarded due to incomplete eye tracking data), found the indented lists to be more effective for searching, while the graphs better supported information processing. Participants completed tasks more quickly using the lists, at least in part due to the more compact visualisation requiring less time to scan. However, they did not improve on accuracy over the use of the graphs. Evidence on which of the two resulted in higher cognitive load was inconclusive, with no statistical difference between the two visualisation techniques. Overall, participants spent more time on information search than in processing; Fu et al. suggest that more research be conducted about building support for search in ontology visualisation.

The paper concludes with pointers to a more detailed investigation of affordances for information search and processing in ontology visualisation tools, including more efficient use of screen space and investigation of the specific requirements of different user types across a wider set of visualisation techniques.

In Aemoo: Linked Data Exploration based on Knowledge Patterns Nuzzolese, Presutti, Gangemi, Peroni & Ciancarini extend previous work on the use of knowledge patterns to counter challenges faced by humans and machines attempting to explore the knowledge content of Linked Data. Nuzzolese et al. argue that encyclopaedic knowledge patterns (EKPs), being derived from a collective intelligence store spanning multiple domains, provide a “cognitively sound” base from which to carry out exploratory discovery of complex, often dynamic, heterogeneous data such as LD.

EKPs are essentially mini-ontologies that define a class, and by mining the links to instances in Wikipedia, retrieve relationships to other classes ranked based on popularity or frequency of use. EKPs map to how humans [contributors to Wikipedia] view (knowledge about) entities and inter-relationships within this knowledge. By retrieving the most important/highly ranked relationships (above a variable threshold), EKPs can be used to summarise knowledge about an entity.

Fig. 2.

Illustrating from left to right, navigation from the EKP for the entity retrieved for the search term “Napoleon” to the EKP corresponding to the “French Invasion of Russia” in “The War of 1812” (see lower snapshot), through an entity of type Military Conflict.

Nuzzolese et al. describe Aemoo (see Fig. 2), a tool targeted at users with neither particular knowledge of the structure or content of a data store, nor expertise in SW technologies. Aemoo demonstrates how EKPs may be used to explore knowledge bases. The aim is to help users focus on the most “important” information about an entity and then navigate to other related information. The authors employ a concept map coupled with a text-based detail window, arguing that mapping the visualisation to the data structure of EKPs aids end-users in forming mental models that support exploration. However, acknowledging potential challenges in the use of the large, complex, unwieldy graphs entailed by the inter-linking of large-scale, multi-dimensional datasets, they restrict their concept maps to a depth one level below the root. Their approach filters out less relevant or important information, thereby also reducing noise. Because Aemoo queries Wikipedia on-the-fly, the information retrieved reflects all updates to the knowledge base. This is in addition to dynamic data retrieved from Twitter and Google news that correspond to the entity in question. Aemoo includes a switch that allows users to retrieve what is described as peculiar knowledge or curiosities – entities linked to the focus that fall below the popularity threshold. This enables users to also visualise information about an entity that would usually not be considered important or relevant.

Fig. 3.

Visualising the results of hierarchical aggregation of Linked Data in the SynopsViz prototype. Clockwise from top, left: dataset statistics; selected classes in a treemap; selected spatial attributes. Bottom, left: middle snapshot zooms in to show detail, before zooming back out and including another temporal attribute.

Nuzzolese et al. present the results of a comparative evaluation, taking Google as a baseline for exploratory search, and RelFinder [35] to test Aemoo’s automatic summarisation against manual filtering and existing approaches to visual querying and exploration. Thirty-two non-technical undergraduates from two universities completed three tasks using Aemoo and, alternately, one of the other two tools. The authors report that each of the three tools performed better than the other two for one task – RelFinder at summarisation, Aemoo at discovering related entities, and Google at discovering relations between entities. Overall, however, Aemoo outperformed both RelFinder and Google search, with RelFinder performing worst. While Aemoo also scored higher for learnability and usability, based on the SUS questionnaire [see 6], there was no significant difference with the other tools.

Based on these evaluation results, Nuzzolese et al. are considering, among others, reusing functionality in RelFinder to present relationships in Aemoo. They conclude with pointers to additional work to improve presentation of, and navigation in, knowledge bases.

Bikakis, Papastefanatos, Skourla & Sellis in A Hierarchical Aggregation Framework for Efficient Multilevel Visual Exploration and Analysis tackle the challenges faced in exploratory analysis of today’s big, heterogeneous, dynamic data, to support immediate and on-going analysis online. Bikakis et al. posit that a hierarchical approach to data processing and aggregation enables customisable construction of data overviews on the fly, from both static and especially dynamic, numerical and temporal data. Followed by support for identifying and drilling into the details of ROIs, the authors aim to provide support for end users in navigating through and making sense of big data.

Their generic model takes as input data regardless of structure and generates output structured to allow the use of varied visualisation techniques for presentation and analysis, hierarchical or not, complemented by statistical summaries of the input data. The authors note also that while they focus on data encoded as RDF, the approach is reusable for data in other forms, provided it is numerical and/or temporal.

Figure 3 shows SynopsViz, a prototype built to demonstrate the results of applying this hierarchical aggregation model to Linked Data. Three visualisation options, a timeline, a treemap and chart views, are provided for visualising the output.

Bikakis et al. assess the effectiveness and efficiency of their approach by comparing it with a baseline described as FLAT – displaying all detail for a dataset. The test measured (initial) response time for data preprocessing and subsequent response time for displaying data of interest, using the DBpedia 2014 dataset [10]. Construction and response time for the two hierarchical approaches presented were roughly equivalent, with minor improvement for the range-based over the content-based tree. For data up to 10,000 triples, FLAT performance approaches that of hierarchical aggregation. Significant cost in construction of the hierarchical approaches was found to be due to communication, notable for small datasets, but decreasing proportionally for larger datasets due to additional time required to render the latter. To improve overall response time, the authors also enable incremental construction for the hierarchical approach. With respect to data properties, the test results show that FLAT was restricted to a maximum of 305,000 triples, while the hierarchical approach successfully processed all data subsets. Further, the latter was found to outperform FLAT for all tests, with increasing difference in response time with dataset size.

Usability evaluation of the SynopsViz tool was also carried out with ten participants with a background in Computer Science. The evaluation involved three sets of tasks across all three approaches to measure users’ ability to (i) discover specific information, (ii) discover data within defined ranges and (iii) compare and rank data subsets. The comparative evaluation was carried out with two datasets, one containing 970 triples and the second almost 38,000. For all three approaches, the authors found significant differences in task completion time for the larger dataset size, due to the much longer time spent navigating through data structure and content. The hierarchical approaches significantly outperformed FLAT for the larger datasets, because data clustering reduced the number of navigation steps required to find matching responses. For the smaller dataset, however, time to completion for FLAT approached that of the hierarchical approach, due mainly to the need for a larger number of navigation steps through the hierarchy. The range task was also particularly difficult to perform using the FLAT approach – cognitive load may have contributed to this as participants had to browse a significant portion of the data (displayed in detail) to complete the task. The range-based and content-based approaches differed slightly in performance time depending on which approach better matched data clustering to task type. The authors note also a key concern of participants that hampered navigation: getting lost during exploration.

The authors conclude with pointers to future work, including a hybrid version that would employ a weighted combination of each of the content-based and range-based approaches to aggregation.

Del Rio, Lebo, Fisher & Salisbury in A Five-Star Rating Scheme to Assess Application Seamlessness assess the cost of coupling multiple visualisation tools together on different platforms, with the aim to promote seamless reuse of visualisation solutions and therefore more effective analysis.

The authors use the munging task in Visual Analytics (VA) – processing of input data to fit a specific tool’s requirements – to illustrate variation in cost due to data structure and semantics, from high-cost 1–3 star mundane data to low-cost, inter-connected, semantically-enriched, 4–5 star data. They extend the PROV ontology [28] to capture data and the analytical process, and to illustrate how inter-connected, semantic, Linked Data may be used to build a bridge between independent datasets and the results of analyses, thus lowering the cost of performing those analyses.

Del Rio et al. describe an application scenario where a student with technological expertise analyses data from multiple stores to obtain information about decommissioned satellites and other equipment abandoned in space. The scenario illustrates the need to switch constantly between semantically-rich data describing the student’s focus, and more mundane representations of this data in visual form, that are easier to interpret but whose content cannot be reused for further analysis. The analysis is therefore constrained, with significant cost incurred by moving between mental models of the data formed as the analysis progresses and alternative representations that break these models. Some of these require further data munging, resulting in attention loss with respect to the data and their semantics. The student then passes on her initial results to a fellow student, who reviews and extends the analysis to confirm her results and answer open questions. The scenario illustrates further difficulty reusing the initial results, due to limited provenance information about the data and the lack of traces about the analytical process followed. Cost again increases when the first student attempts to reuse the analysis results of her colleague; while the new visualisations provide further information, they are not directly connected to the underlying linked and semantically-enriched data.

Del Rio et al. utilise the outcomes of this scenario to derive an application seamlessness metric that takes into account the initial cost of analysis and the cost of reusing the results of prior analyses. The metric is defined using ontology restrictions that assign costs ranging from one to five stars. The costs reflect the impact of employing a set of applications or following specific approaches in the data analysis process. The proposal harnesses the strengths of both the VA and SW communities for managing data to avoid pain points and reduce cost in data analysis, instead increasing effectiveness. A key component of the approach is maintaining the links between the analyst, the analytical tools in use, and input data (including previous analysis results).

The authors conclude with proposals for future work to take into account also analysts’ expertise and background in measuring the cost of using selected tools or approaches. They also discuss refining their guidelines to capture more effectively the cost difference between individual star ratings, as opposed to ranges, in their seamlessness metric.

Today’s inter-connected, data-driven economy requires decision-makers to consider the impact of multiple external factors using data from a variety of sources and increasingly outside their domain, in order to feed wider contextual data into informed decision-making. However, the cost of access to third party data, especially where proprietary, and support for effective analysis of the resulting heterogeneous datasets, are often beyond the means of smaller organisations and research institutions. Brasoveanu, Sabou, Scharl, Hubmann-Haidvogel & Fischl in Visualizing Statistical Linked Knowledge Sources for Decision Support look at the potential of the growing collection of statistical Linked Open Datasets to fill this need. These datasets, including the Eurostat database [11] and the World Bank Catalog [41] are built on the RDF Data Cube Vocabulary [8].

Brasoveanu et al. identify a set of challenges inherent to the exploration and analysis of large-scale, heterogeneous, inter-connected data, made more complex by the unique challenges found in statistical LD analysis – including variation in data structure and content, inconsistent and poor use of data schemas, and frequent failures of dynamic data portals such as SPARQL endpoints. They therefore highlight the need for integrated visual, multi-perspective analysis solutions that enable complex question-answering over such data, that scale, and that are reusable beyond specific use cases.

The authors describe the iterative, scenario-driven process they followed through requirements collection, design and prototyping, following best practices and guidelines for visual analysis, with a focus on the specific requirements of Linked Data and statistical data. This process resulted in a dashboard containing multiple, coordinated components for data preprocessing and integration, visualisation-driven analysis, and for effective presentation of the often also multi-dimensional analysis results. Based on their findings – challenges faced and successful application of research – they propose a set of guidelines and workflows for tackling the challenges in building decision support systems over Linked Statistical Data.

Brasoveanu et al. illustrate the advantages in providing multiple perspectives on the data (slices) and the results of analyses using a set of scenarios about decision support in the tourism and telecommunications industries. They then summarise the results of exploratory evaluation of their ETIHQ dashboard, to measure usability against traditional approaches to cross-domain analysis for decision support, and identify additional functionality required by target users. Sixteen participants took part in the evaluation, all in their usual work environment: ten tourism experts based in different countries and six researchers. They completed two activities with the dashboard, comprising preset and user-defined tasks, and a third activity using their usual working tools. The ETIHQ dashboard was observed to perform significantly better than manual (traditional) approaches to question-answering, with participants providing more detailed and precise responses, accompanied by a gain in task completion time of almost 30%. Recommendations for improvements to the dashboard included, among others, the use of a simpler workflow, the implementation of search features, and the use of domain terminology (as opposed to terminology reflecting the use of SW technology for encoding data). Participants also requested the inclusion of news and social media data.

The authors conclude with a description of updates to the dashboard based on feedback from the usability evaluation.

We summarise in Table 2 the contributions in this special issue by looking at the set of visual presentation and analysis features that was also used in Table 1 to compare other related work. For completeness, two additional features relevant to LD visualisation and analysis are included: faceted search/browse and support for editing underlying data. Neither of these features was implemented in any of the tools in Table 1, and we only see limited implementation here.

5.Future challenges for Linked Data visualisation

In their introduction to the Semantic Web Journal’s first set of LD description papers in 2013, Hitzler and Janowicz [21] identify Linked Data as “enablers in research”, stating its importance. We reiterate this statement confidently: even if many challenges remain, LD has significant value as a rich, structured source of knowledge for both research and practical applications.

The challenges discussed in this special issue revolve mainly around making LD accessible to end-users, primarily through interactive, visual representations of this novel form of data whose characteristics are quite unique, as discussed earlier. While these characteristics are what make LD such enablers, they are also the reason why it is proving so difficult to design effective user interfaces for LD and for the Web of Data at large.

As demonstrated in the papers in this special issue, we continue to see advances in research on the use of visualisation for making LD more accessible. Efforts have been moving from producing simple, basic demonstrators of LD browsing to more elaborate tools that integrate data capture, exploration, analysis and/or presentation in a visualisation-driven front-end. However, there is still significant room for improvement.

Hendler [ 20] in 2008 raised the “chicken-and-egg problem” regarding the adoption of Semantic Web technology outside the community. The technology was not sufficiently proven at the time, even though RDF and SPARQL had been introduced several years before, and examples of success using ontologies to describe data already existed. Government, state-run institutions and industry were reluctant to dedicate resources to encoding and sharing their varied data to grow this new “Web of Data”, but data was required to feed into the development of sophisticated tools for managing and interacting with this new Web. Seven years later, Neish [40] describes several examples of successful use of Linked Data in different domains outside the Semantic Web community, in government, education and industry. He stresses, however, that links across independently created datasets rarely span across domains. Further, technological barriers continue to hamper the adoption of LD principles in data generation and sharing, with a corresponding increase in economic and other risks following these principles for data management.

User interface design for LD is still in its infancy. This is reflected in publications describing design and prototypes developed within the research community for LD visualisation, which are targeted predominantly at workshops rather than journals or research tracks in conferences. However, workshops dedicated to discussion of research on interactive visualisation tools for LD are one avenue that will help foster research as the generation and consumption of LD increases, and as LD use spreads over more and more application areas. Interestingly, Neish [40] concludes that incremental adoption of LD may help avoid the acknowledged “chicken-and-egg problem”, by ensuring that the technology is well developed and proven before the Linked Data initiative starts to play a key role in helping to resolve newer challenges such as those posed by today’s complex, distributed, heterogeneous, big data [see also 39,58].