The number and the size of Linked Data sources are constantly increasing. In some lucky case, the data source is equipped with a tool that guides and helps the user during the exploration of the data, but in most cases, the data are published as an RDF dump through a SPARQL endpoint that can be accessed only through SPARQL queries. Although the RDF format was designed to be processed by machines, there is a strong need for visualization and exploration tools. Data visualizations make big and small linked data easier for the human brain to understand, and visualization also makes it easier to detect patterns, trends, and outliers in groups of data. For this reason, we developed a tool called H-BOLD (Highlevel Visualization over Big Linked Open Data). H-BOLD aims to help the user exploring the content of a Linked Data by providing a high-level view of the structure of the dataset and an interactive exploration that allows users to focus on the connections and attributes of one or more classes. Moreover, it provides a visual interface for querying the endpoint that automatically generates SPARQL queries.
Since 2006, the year in which sir Tim Berners-Lee coined the
term Linked Data (LD), leading to a new way in which data can
be structured and accessed through the Internet, the number
of LD exploded. Starting from governments, many institutions,
enterprises, and privates adopted this method for publishing data.
As a result, there is a huge number of LD that can be accessed
through the Internet. Unfortunately, endpoints are big containers
of triples. They can contain every kind of information and the
high number of triples required to express a concept made LD
visualization a non-trivial task. Due to the volume and the variety
of information, it is hard to find a common procedure for
exploring every dataset but nonetheless, several research groups tried
to address this task [
In this paper, we present a new version of the tool H-BOLD (High-level visualizations on Big Open Linked Data). H-BOLD enables the exploratory search and multilevel analysis of Big 2
H-BOLD1 (High-level visualizations on Big Open Linked Data) is
a tool, available online, for visualizing, and interacting with LD.
It is defined in the context of hierarchical and interactive visual
exploration and analysis over LD [
H-BOLD starts from our past experience with the tool LODeX,
a tool able to automatically provide a summarization of a LD,
including its inferred schema [
H-BOLD is composed of a server layer and a presentation layer that will be described in the following. 2.1
The server layer aims to generate high-level representations of a
set of LD starting from a list of SPARQL endpoints. The outputs
are the Cluster Schema (a high level representation of a complex
and big source) and the Schema Summary (a low level
representation of the instanced classes within a source). The SPARQL
endpoint list is created starting from the old list of endpoints used
in LODeX and adding new SPARQL endpoints that are present
on Open data portals. Moreover, we enable users to manually
add a new URL for a SPARQL endpoint they wish to visualize
and explore in H-BOLD. On each endpoint a set of queries is
executed in order to extract structural and statistical information
that describe the LD, this phase is called Index Extraction. In
particular, the indexes are the number of instances, the number
of classes, the list of classes with the respective properties and
the number of instances belonging to a specific class. The Index
Extraction is able to deal with the performance issues of the
different implementations of SPARQL endpoints by using pattern
strategies [
Starting from the indexes is then possible to create a Schema
Summary of the LD, a pseudograph that represents, through
nodes and arches, the relations between the various instantiated
classes of the dataset [
we are dealing with big sources, it happens that the number
of classes is high, thus the graph contains a high number of
nodes and the visualization results complex and confused. On
the Schema Summary, a set of community detection techniques
has been used to create a high-level visualization for Big LD
[
The Schema Summary and Cluster Schema ofer several advantages: they can be easily memorized and retrieved on the MongoDB improving data recovery performance and graph visualization. 2.2
In the presentation layer, the first step for a user is the selection of a dataset, then, the user can start the exploration of the Cluster Schema or the exploration of the Schema Summary. The first is more concise, while the second is the complete visualization of the structural information on the LD.
If the user opts for the Cluster Schema (see Figure 2, steps 1), he/she will see a shrank representation of the Schema Summary that is obtained by applying a community detection algorithm. From the Cluster Schema, by selecting a class within a cluster, a new visualization focused on the selected class is proposed to the user. The user might then further explore the class, its connections with other classes and its attributes, or can iteratively increase the graph displayed by expanding the connections starting from some classes (nodes in the graph). In each partial representation of the Schema Summary, the user is informed about the percentage of the instances represented by the graph and the total number of nodes (see Figure 2, steps 2 and 3). This expansion can be repeated until all the classes are displayed, as in the Schema Summary visualization.
Another option for the user to start is from the exploration of the Schema Summary, here, he/she will see a complete graph containing all the instantiated classes of the LD (see Figure 2, step 4). The user can focus on a particular class and explore its attributes and properties. Figure 2 shows the visualization steps in H-BOLD starting from the Cluster Schema and then selecting a class and expanding the graph till the visualization of the entire Schema Summary over the Scholarly LD2. 3
The H-BOLD application has been renovated both in the server layer and in the presentation layer. The server layer is implemented entirely in python and it uses MongoDB for the storage of information. With respect to the previous version of H-BOLD, it has been equipped with new features for making the application more responsive to the user requests. First of all, to enhance the understanding of the behavior of the tool, a guideline has been drawn up: the application code has been studied and various flowcharts have been extracted. A process for continuously updating and enriching our collection of SPARQL endpoints has been implemented (see Subsection 3.1 for further details). The Cluster Schema creation has been re-implemented and included in the Server Layer, while previously was calculated on-the-fly in the Presentation Layer (Subsection 3.2). The list of SPARQL endpoints has been enriched thanks to the integration of three new open data portals (Subsection 3.3).
The presentation layer went through a complete re-engineerization. Due to many incompatibilities between the current browsers and the version of Polymer3 in use (i.e. a javascript framework developed by Google that helps building web application through the concept of web components), we were forced to re-implement the presentation layer of H-BOLD.
We started searching alternatives that could replace Polymer. At the end of this research, we decided to move towards Bootstrap4, a javascript framework developed by Twitter for building and styling web applications. Therefore, the interface of the web application has been completed rebuilt with Bootstrap. The
javascript graphic library, D35, already in use in H-BOLD, has been integrated with Bootstrap to create interactive visualizations. Moreover, new features ameliorate H-BOLD. A new interface for allowing user to manually insert a SPARQL endpoint has been added (Subsection 3.4). New visualization layouts have been conceived for the exploration of the Cluster Schema and the Schema Summary (Subsection 3.5). 3.1
Adding new content inside a SPARQL endpoint is a simple task since the only constraint is that data must be represented in one of the RDF serialization formats. Adding new classes or relations is just as simple as adding new instances. As a consequence, the structure and also the content of a LD could change very often. In H-BOLD, we want to display the most updated version of the dataset we indexed. For this reason, we automated the procedure of indexes extraction and Schema Summary and Cluster Schema generation to run daily. Working with SPARQL Endpoints since 2014, we noticed two important aspects. First, a SPARQL Endpoint might be often not available6, but this does not mean that it is completely out of order, it might work again after 1 or 2 days. Second, LD do not change daily, they usually changed weekly, or monthly, or do not change ever. For these reasons, it is useless to run the index extraction over all the datasets daily, it is enough to run it weekly. However, since a SPARQL endpoint might not be available one day, and maybe be online the next day, we should check its availability. Therefore, we decided to store the date of the last index extraction for each SPARQL endpoint. If the last index extraction was executed more than seven days before, then we do not update the information for that LD, however, if some things went wrong with the last index extraction because the endpoint was not available, we can repeat the index extraction every day.
3.2
The Cluster Schema is the core part that has been inserted in
H-BOLD, starting from the predecessor LODeX. The Cluster
Schema allows to visualize in an aggregated way all the classes of
a LD. In the previous demo of H-BOLD [
In the previous version of H-BOLD, the list of datasets that users could visualize was manually created from a list of SPARQL endpoints available on DataHub. Unfortunately, the endpoint we queried for obtaining such information is no more available but the same data can be found at this link 7. Moreover, that list was quite old and some of the endpoints were no longer available. For this reason, we conducted a research to identify which were the most relevant open data portals that contains links to SPARQL endpoints. Then, we extended both the server
layer and the presentation layer. The server layer was set up to query three new open data portals. In particular, now, H-BOLD search for new endpoints in the following portals: • European Data Portal8. It contains open data published over the diferent european countries, regions and local administrations open data portals; • EU Open Data Portal9. It holds the data produced by the diferent organization of the European Union; • IO Data Science of Paris10. Born with the idea of creating new synergies between data, it contains a metadata description of several LD.
The portals were deeply explored in order to gain knowledge over their content. Initially, we tried to produce three ad-hoc queries for extracting the maximum amount possible of SPARQL endpoints but then we found that the query presented in Listing 1 perfectly fits all the portals.
Listing 1: Query sent to the open data portals to extract a list of SPARQL endpoints PREFIX d c a t : < h t t p : / / www. w3 . org / ns / d c a t # > PREFIX dc : < h t t p : / / p u r l . org / dc / t e r m s / > SELECT ? d a t a s e t ? t i t l e ? u r l WHERE { ? d a t a s e t a d c a t : D a t a s e t . ? d a t a s e t dc : t i t l e ? t i t l e . ? d a t a s e t d c a t : d i s t r i b u t i o n ? d i s t r i b u t i o n . ? d i s t r i b u t i o n d c a t : a c c e s s U R L ? u r l .
f i l t e r ( regex ( ? u r l , ' s p a r q l ' ) ) . }
With this research, we discovered 65 SPARQL endpoints on the European Data Portal, 9 SPARQL endpoints on the EU Open Data Portal and 15 SPARQL endpoints on the IO Data Science of Paris. Some endpoints were already present in H-BOLD, therefore we were efectively able to increment the number of endpoints in our collection by 70 units. As a result, the number of endpoints listed in H-BOLD raised from 610 to 680. Since some of them are not working or are not compatible with the index extraction phase of H-BOLD, we were able to index and expose the structure of 20 new datasets raising the number of indexed endpoints from 110 to 130.
3.4
Even with the crawling of SPARQL endpoint from open data
portals, we are not able to reach, index and expose every SPARQL
endpoint available on internet. Therefore, to further increase
the number of indexed datasets, we integrated in H-BOLD a
procedure through which the user is able to upload the URL of a
SPARQL endpoint and to see, this new dataset listed among the
others in the H-BOLD dataset list (we tested this procedure over
some open datasets [
The previous version of H-BOLD adopted only graphs for
displaying the information (as shown in Figure 2) and in some occasion
they were not suited for extracting the maximum of the
information. For instance, the number of instances belonging to a
class is hard to extrapolate from a graph and it is also hard to
understand which classes have been aggregate in what cluster.
Moreover, when the number of classes and relations is high, it is
hard to understand the various connection. For this reason we
implemented four supplementary visualization layouts, three for
the Cluster Schema and one for the Schema Summary. We drew
inspiration from an extensive analysis over other LD
visualization tools [
3.5.1 Treemap visualizations of the Cluster Schema. Treemaps are an alternative way of visualising the hierarchical structure of a Cluster Schema while also displaying quantities for each cluster and each class within the cluster via area size. Each cluster is assigned to a rectangle area with a specific color and their classes rectangles nested inside of it. When a quantity is assigned to a class, its rectangle area size is displayed in proportion to that quantity and to the other quantities within the same cluster in a part-to-whole relationship. Also, the area size of the cluster is the total of its classes. If no quantity is assigned to a class, then its area is divided equally amongst the other classes within its cluster. The treemap built over the instance’s count, shown in Figure 4), highlights the classes with the higher number of instances, the size of the clusters and the predominance of some classes in terms of instances. 3.5.2 Sunburst visualizations of the Cluster Schema. The Sunburst Chart visualization (Figure 5) shows the hierarchy through a series of rings, that is sliced for each category node. The inner ring represents the clusters while the outer ring shows the classes grouped by the clusters.
3.5.3 Circle Pack visualizations of the Cluster Schema. The Circle Packing (Figure 6) is a variation of a Treemap that uses circles instead of rectangles. Containment within each circle represents a level in the hierarchy: each branch of the tree is represented as a circle and its sub-branches are represented as circles inside it. Similarly to the rectangles in the Treemap, the circles might have diferent dimensions. As displayed in Figure 6, the inner circles represent the classes, while the intermediate circles represent the clusters, an external circle represents the entire dataset. In some cases, a cluster can contain only one class.
3.5.4 Hierarchical edge bundling visualization for the Schema
Summary. Hierarchical edge bundling is a method developed by
Holten in 2006 [
Using the Hierarchical Edge Bundling layout, the classes are displayed over an invisible circumference and the properties are arcs within the circumference. This layout is perfectly suited for understanding links within the classes and the domain and range of the properties that connect the classes. As an example, in Figure 7, the node in bold (Event) is the class of interest, the node in green (Situation) is the rdfs:Range class of the properties that connects it to the class of interest and the nodes in red (Vevent, SessionEvent, ConferenceSeries and InformationObject) are the rdfs:Domain classes of the properties that connect them to the class of interest. 4
In [
In [
In this paper, we presented H-BOLD, a tool for multilevel interactive visual exploration of Big LD that has been enhanced with new visualizations: TreeMap, Sunburst Chart and Circle Packing. H-BOLD has been tested on 130 Big LD showing good performances.
In the next future, we intend to raise the number of endpoints indexed in H-BOLD by improving the index extraction procedure and by querying new repositories that collect SPARQL endpoints metadata. Moreover, we intend to evaluate the efectiveness of HBOLD as a visualization tool through a survey involving diferent kinds of LD consumers: practitioners, unskilled users, domain experts. 6
This work has been partially supported by the TRAFAIR project 2017-EU-IA-0167, co-financed by the Connecting Europe Facility of the European Union and the “Networking over Linked Data" of the University of Modena and Reggio Emilia within FAR2019. The contents of this publication are the sole responsibility of its authors and do not necessarily reflect the opinion of the European