Over the last decades, cultural heritage institutions have provided extensive machine-readable data, such as bibliographic and archival metadata, full-text collections, and authority records containing multitudes of implicit and explicit statements about the social relations between various types of entities. In this paper, we discuss how approaches to the creation and operation of advanced research infrastructure for historical network analysis (HNA) based on heterogeneous data sources from cultural heritage institutions can be examined and evaluated. Based on our interdisciplinary research, we describe challenges and strategies with a special focus on the issue of data processing, sketch out the advantages of human-centered project design in the form of a preliminary co-design workshop, and present an iterative approach to data visualization.
The study of historical events is relevant to many disciplines in the digital
humanities, with the analysis of relationships between agents often being
crucial for the understanding and explanation of social, political, and cultural
phenomena. Given that historical research is heavily dependent on
information from the respective time period, the combination of as many historical
sources as possible is essential for the reconstruction of historical networks
– and this is where the method of historical network analysis (HNA) comes
into play. Derived from social network analysis, HNA is characterized by the
same dependency on numerous historical sources that ideally support each
other
One limiting factor in HNA can be a lack of awareness with regard to the availability of suitable research data. At the same time, over the past decades, cultural heritage institutions have produced very large amounts of machine-readable and, in many cases, standardized and well-organized data in the form of bibliographic and archival metadata, full-text collections, and sets of authority or reference records. These datasets contain a plethora of implicit and explicit statements about social relations, which can in turn be exploited for HNA research. However, systematically combining multiple data sources (not to mention extracting and visualizing the complex resulting networks) currently requires extensive knowledge in graph theory as well as time-consuming manual work carried out by the individual researcher. One reason for this is the heterogeneity of the data sources made available by cultural heritage institutions, for example, in terms of data formats.
The research project SoNAR (IDH): Interfaces to Data for Historical Social Network Analysis and Research1 addresses this issue. We examine and evaluate approaches to the development and operation of HNA-supporting research infrastructure based on heterogeneous cultural heritage data. In this paper, we present a number of preliminary insights related to the process of modeling and transforming heterogeneous data sources, and to the design of user-centered visualization for historical social networks. By sharing our approach and its accompanying challenges, we aim to contribute to the ongoing discussion on the suitability of bibliographic big data for HNA and the development of corresponding research technologies. 2
The following section gives an overview of previous research, and discusses projects related to graph modeling and visualization approaches within the digital humanities from the perspective of historical network analysis. 2.1
In recent years, open knowledge graphs have frequently been used as an
alternative to a document-based approach
Even though these initiatives provide, among other things, access to new, previously unidentifiable or implicit information, they do not primarily focus on the extraction of network data. Therefore, HNA researchers are often left to create their own individual graphs after gathering data that is suitable to address their research question(s), in many cases using open source software tools such as Gephi,8 Palladio,9 or VennMaker,10.
Along with the establishment of network analysis as a method in historical research, there has been an increase in joint research projects that are focused on the extraction of historical networks within the social sciences and the humanities.
For example, the project Six Degrees of Francis Bacon11 applies statistical
methods to the base data with the goal of inferring relations that permit the
reconstruction and visualization of historical social networks in Early
Modern Britain. The project allows for the expansion and curation of the data
through collaborative annotation by the users
These are only a few examples of the ongoing eofrts to provide users with direct access to networks in existing data collections. In our project, we are working with data sources that have not been modeled for HNA before. Our generic data approach is closely connected to similar projects, like the North American cooperative SNAC – Social Networks in Archival Context15 and the French project PIAAF,16, which both have a strong focus on archival metadata and full texts. 2.2
As far as the visualization of data for HNA is concerned, many interfaces
have been developed over the years that oefr explorative, web-based
network visualization tools for historical network analysis. Examples include
the above-mentioned Six Degrees of Francis Bacon
Graph visualization is an extensive field in itself, which is accompanied
by a substantial body of literature on issues such as graph-related algorithms
A key element in the examination and development of a new research
infrastructure designed for human-computer interaction is how well it meets the
needs of the people it is intended to assist. This human-centered approach
is closely related to Grounded Theory, which generates inductive results by
means of sociological methods
The interdisciplinary project SoNAR (IDH), which studies the potential of large heterogeneous data collections for HNA, includes partners from the ifelds of historiography, information visualization, and artificial intelligence, as well as computer and information science. This variety of disciplines opens diefrent perspectives on the requirements and challenges connected to the use of heterogeneous (meta)data for HNA. What distinguishes our approach is the synchronous operation of all components of the project, so that the design of the data technology, the development of a model research design for HNA, and the development of innovative visualization and interface approaches with the involvement of HNA experts are all intertwined and influence one another.
The project is based on heterogeneous source data from authority files,
bibliographic records, and full texts. The data is available in various
XMLbased formats such as MARC21
Since the source data – describing entities (authority files) and resources (bibliographic files) – is encoded in various formats, these formats must first be analyzed in order to enable the design of an appropriate data model and allow their transformation into a uniform, generic format. Full texts are prepared for automatic enrichment (i. e. named entity recognition and linking) and converted to a corresponding format. 4
In this section, we will give an overview of the data transformation and graph modeling process, and outline the challenges that we have encountered along the way.
The technical goal of our project is the integration of the various source
datasets into a common research infrastructure. We currently use the graph
database Neo4j,26 which is well suited to the eficient storage and
highperformance analysis of large amounts of highly networked information
There are a total of 9 entity types extracted from the source data: 1. Person PerName; 2. Corporate body CorpName; 3. Place or geographic name GeoName; 4. Conference or event MeetName; 5. Subject heading TopicTerm; 6. Work UniTitle; 7. Temporal information ChronTerm; 8. Information about ISIL27 IsilTerm; 9. Resource Resource. 23https://kalliope-verbund.info/en/index.html 24http://zefys.staatsbibliothek-berlin.de/index.php?id=start&L=1 25https://www.dnb.de/EN/Sammlungen/DEA/Exilpresse/exilpresse_node.html 26https://neo4j.com 27International Standard Identifier for Libraries and Related Organizations
Six entity types (i. e., person PerName, corporate body CorpName, place or geographic name GeoName, conference or event MeetName, subject heading TopicTerm, and work UniTitle) are taken from the corresponding classes of the authority ifles. Bibliographic entities are represented as Resource. We added two types to this list: ChronTerm, which describes temporal information encoded in entity types from authority files; and IsilTerm, which is used to identify the libraries related to other entity types. Each entity has general features, such as a unique source identifier, URI, name, link, etc., and specific features, such as age, gender, coordinates, etc. Furthermore, there are also nine relation types that correspond to entity types, such as RelationToPerName, RelationToCorpName, RelationToGeoName. Relations between entities include information about the relation source, relation source type, information about temporal validity, and additional information (Figure 1). 4.1
While the relations between entities are explicitly described in authority files, relations between actors such as persons or corporate bodies that are identiifed or defined in the resource are only implicitly encoded in bibliographic ifles. Our aim is to automatically infer these implicit relations with the assistance of a set of strict guidelines (e. g. a connection between two persons can be assumed if both are co-authors of a scientific publication), and to make them available as explicitly encoded data. In order to derive corresponding relation types, the role of actors regarding a specific resource (e. g. as author, editor, or addressee) and the resource type (bibliographic files of primary sources of the Kalliope Union Catalog and of secondary sources of the German National Library and the German Union Catalog of Serials) are to be taken into account. Using this approach, we were able to infer additional relations (but marked them as computed), for instance between co-authors, co-publishers, and authors/addressees, to further enrich the data.
In order to prepare full texts for analysis, named entities are automatically recognized, disambiguated, and linked to their associated authority files (e. g. the Integrated Authority File or Wikidata28). Next, relations between detected entities are automatically recognized, added to the graph database, and connected with their respective full texts, represented as nodes. 4.2
Overall, the authority and bibliographic files used by us contain approximately 30 million records that describe entities and resources in detail. As was to be expected, normalization of the data revealed a number of errors and inconsistencies. In this section, we would like to describe some particularly problematic areas in more detail and suggest possible solutions.
We have modeled and transformed data for the graph database in such a way that identifiers are used as coordinates for relations between entities. In the Integrated Authority File, entities with old identifiers were found, so that an appropriate connection of two entities was not possible. The first challenge was to detect old identifiers and replace them with valid ones in order to enable error-free representation. All replacements were written in a log file. However, during a consistency check we also found relations to entities within the source data that were without identifiers. Since such entities could not be clearly assigned to existing entities with identifiers, ambiguous relations of this kind had to be ignored.
Information that was encrypted in internal codes in the Integrated Authority File, the German National Library, and the German Union Catalogue of Serials (in format MARC21) was also checked for codes of general and specific entity types, codes of relation types between an agent and a resource, and country codes. Further examinations were performed on the consistency of entity names, resource titles, and identifiers. Again, all errors or inconsistencies were written in a log file.
Building on the conclusions that we were able to draw from testing Neo4j, we decided to adapt the data model to our needs. In order to simplify searching and filtering according to temporal dimension, time information from the source data was adjusted. First, while retaining the source data, we additionally separated time intervals, noted as “begin” and “end.” Second, in order to facilitate more performant visualization and querying of the data, we added a feature to resource descriptions that reflects the year of publication (in addition to the publication date). Thirdly, diefring time expressions in MARC21 and EAD were normalized.
We also decided to change gender-specific names of professions. These are represented in the Integrated Authority File as two diefrent entities with their own identifiers, male and female. Conceptually, however, what we are dealing with is a single entity with two versions, so these versions must be merged in the graph database and represented as a node. One challenge is to adequately display all information from the two versions without making the search more dificult. In this case, we are currently still looking for a suitable solution. 5
In accordance with the principles of grounded evaluation
At the beginning of the project, it was important to us to stimulate
discussions on the potential of bibliographic (meta)data for HNA, and on the
requirements for the visualization of historical networks. Following the
approach proposed by
The workshop consisted of ten participants, including four historical/social network practitioners as domain experts, two project-internal information visualization designers/engineers, two members of our project-internal evaluation team, one member of our team of data scientists (responsible for the data transformation), and an external participant who had a background in design and previous experience with the co-design format. The interdisciplinary composition of the group was intended to enrich the discussion by oefring a multitude of perspectives on the topic of HNA through the lens of HNA experts, with fresh insights being provided by participants from other (project-relevant) fields. Since we aim to develop an infrastructure for HNA that can be used by researchers from all disciplines working with this method, the participation of experts from fields other than history was especially welcome.
The workshop was scheduled for three hours in total. As suggested by
We started the process of conceptualizing network visualizations with a
short, hands-on visualization exercise, during which the participants were
asked to visualize a very small social network (ten nodes) based on a data
matrix we provided. After this warm-up, we gave a short introduction to the
goals of our project and the data we are using. The participants were then
asked to create a collage depicting possible approaches to HNA research with
our specific data and project in mind (see Figure 2). For the collages, we
supplied a variety of materials (e. g. construction paper, pencils and
markers, sticky notes). While
After about 30 minutes, each of the collages was discussed. First, the participants not involved in making a collage were asked to interpret and speculate about what they were seeing. Afterwards, the creators of the collages were asked to give explanations and discuss their approach with the group. In this step, the almost inevitable misinterpretations were meant to foster further discussion and novel ideas. In the final step, each participant was asked to give a closing statement recapitulating the most important insights from the process and the most prominent topics or themes in the discussion.
For further analysis and documentation, the entire workshop was audiorecorded and photographed. The audio recordings were transcribed and encoded in a tool for qualitative data analysis. This allowed us to assess various qualitative aspects of the workshop discussions at a later date. As pointed out above, the goal of our workshop was not to create functional wireframes or concrete interaction principles, but to stimulate discussion, foster sensibility towards the domain and data, and highlight important domain-specific research aspects and challenges. The following section will discuss some of the most relevant insights concerning our visualization process. 5.2
We noticed two diefrent types of statement. On a more abstract level, the participants expressed various information needs that commonly arise in the process of their research. In some cases, however, the conversation and the collages yielded very concrete ideas regarding possible features of an HNA infrastructure that would address these needs. As mentioned before, the latter were not regarded as direct assignments to be fulfilled in the visualization process, but rather as indicators for the participant’s general receptiveness towards various properties of the user interface of an HNA infrastructure. Table 1 and 2 summarize the main aspects of the workshop discussions in the form of needs and features.
The most pressing topic in the discussions was the envisioned user approaches and use cases the infrastructure is expected to support. Seven of the ten participants expressed the hope that the visualizations could generate new perspectives, thereby creating forms of access to the data that would hardly be available based on non-machine-supported cognitive work. In this
Uncertainty Data potential Graph density Entry points Data explanations 30 25 27 26 17 16 7 7 4 4 4 4
Tie metrics Other filters Export and citation Location filter Source linking 22 18 16 13 8 6 4 4 3 4 3 4 context, one participant explicitly emphasized the potential of visualizations to raise new questions:
What kind of relationships you are looking for in the data is something you often notice in the very moment that you look at the pile for the ifrst time. 29
Since the participants were aware of the fact that we are confronted with a very large amount of data, which can hardly be presented in its entirety (see Section 3), a discussion of possible entry points emerged. There was consensus on the importance of filter options, most importantly time filters: Without timelines, the visualizations are of no use to me – neither for analysis, nor for the presentation of results.
In addition to timelines, other filters (e. g. node type and node source) were considered a prerequisite for data exploration. Three participants also mentioned the importance of location filters (e. g. through a map view).
Participants with more HNA experience explicitly stressed the essential role of a multi-layered approach. The capacity to display the evolution of relations (e. g. through time and location) was described as the distinctive feature of HNA when compared to the non-historical analysis of social networks. The sole option of static display was considered insuficient.
Along with possible entry points, another important topic raised in the discussion was data complexity, with introductions and explanations regarding the underlying data being identified as particularly crucial. Some participants suggested addressing this issue with the help of concrete use cases that could give potential users a more specific idea of the possibilities aofrded by the HNA infrastructure.
29All quotes translated from German into English.
About half of the participants cited the ability to quantify network characteristics as graph metrics during the research process as one of their main motivations for using HNA methods. This includes indicators such as the clustering coeficient, closeness centrality, degree distribution, degree centrality, and betweenness centrality. Four participants also considered density within a selected sample of nodes to be a relevant indicator for a given dataset’s potential for network analysis. After the first cluster of possible approaches had been discussed, one participant highlighted the added value of graph metrics when it comes to the identification of anomalies in the data: What all these things are actually about is that we are looking for patterns!
Some participants also stressed the potential of tie metrics to accommodate a variety of relation types, and expressed the desire to have the weight of edge properties visualized:
It is of course a big diefrence whether you are a family member […] or whether you are a correspondence partner or whether you met at a congress during a coefe break. These are all relationships, but of course they have diefrent weights in their interpretation. This is, for example, something we would like to see in the visualization.
This statement is representative of another central topic discussed in the workshop, namely the visual marking of missing or uncertain information in the data which can, for example, be the result of inconsistencies in the metadata fields (see Section 4.2). The design expert considered this to be a major desideratum:
I think this is not done enough in current visualizations to show uncertainties of data.
With regard to the scientific standards of HNA research, a final major issue was the export and citation of the visualizations. This, of course, requires unambiguous and persistent provenance links to the source of each data point, as well as timestamps of the corresponding data import.
Many of the results of our co-design workshop match the challenges in
information visualization discussed in the pertinent literature. In the
following section, we will draw on these results to describe our prototyping
approach and process.
The dataset of our project is comprised of a number of elements that go well
beyond what can be perceptually or cognitively grasped at a glance. When
it comes to encoding, for example, the sheer amount of nodes and relations
poses technological as well as visual challenges
Even though our workshop, our conversations with domain experts, and
existing task taxonomies
Compared to the potentially shortest path to a finished ‘tool,’ our method
resembles a curiosity-driven ‘sandcastling’
From the beginning, the entire project has been conducted in an interdisciplinary and concurrent mode, without any delays between its individual steps; data processing, case study development, visualization, and evaluation all occur alongside one another. Initially, the data was neither processed for visualization, nor was it accessible via some form of API, which only allowed us to work with small subsets of selected data. While this made it dificult to anticipate all of the facets and challenges associated with handling the full extent of the data, working with data subsets early on gave us the ability to exert iterative influence on the data processing and the data model.
Instead of trying to combine all potential features and ideas into a single
prototype, our approach focuses on small, separate problems and ideas
through a multitude of rough prototypes. Many of our design studies or
prototypes have been developed in close collaboration with our own HNA
specialists, and/or draw extensively on input from our workshop or other
external sources of expertise, whereas others are more experimental in nature,
and are often the product of spontaneous impulses. For the most part, the
following examples were designed with the data visualization library D3.js
Figure 3, for instance, shows two small design studies from the beginning of the project, without using real data: the one to the left is a visualization of levels of relation uncertainty, while the one to the right represents the testing of an interaction concept with the goal of reducing complexity by merging multiple edges and allowing users to fan them out on demand.
As an example of the influence exerted by visualization on the data model, an early prototype which clusters persons in a small subset of the data based on related topic terms – in most cases occupational titles – revealed that these titles are frequently gendered30 in our base data (GND), which means that men and women are often not related to the same topic term, even though they practice the same profession. This unexpected diefrentiation in the data is highly relevant when it comes to search queries and visualization, since it is quite possible that some researchers do not diefrentiate by gender, and only use the male form that was traditionally considered to be generic. One eefct of this diefrentiation in the data can be seen in another interactive prototype (see Figure 4), where it is possible to select a specific year in the data with a slider, visualizing top topic terms related to persons who were alive in the selected timespan (female-gendered topic terms are colored in orange). The goal of this prototype was to explore the potential of overviews to reveal aspects of the data that might, at a later point, act as entry points 30Many German occupational titles are gendered and exist in a male and a female form, as with the English ‘actor’ and ‘actress.’ for specific search interests.
Another experimental prototype (see Figure 5) of a small subset of our
data also focuses on topic terms and the temporality of the data; an aspect
frequently highlighted as important by some of our HNA experts in the
workshop. Here, the dimensionality reduction technique UMAP
In some cases, as with Figure 6, we developed prototypes out of curiosity for very specific research questions, for example: “Are network communities in the data subset mostly composed of contemporary nodes or do communities stretch over multiple generations?” Here, the prototyping process allowed us to test specific algorithm implementations and design strategies, while at the same time being able to obtain deeper insights into the data.
While our research is still in progress, the experiences mentioned above illustrate the benefits of staying curious and open to experimentation throughout the analysis and visualization process. Even though many ideas and concepts are inspired by existing research in the field and, of course, the expertise of our domain specialists, we see additional value in experimenting with the data and generating a multitude of visual representations, even if this means knowingly taking detours. It is precisely these more experimental pathways that can lead to new ideas for tools, or generate fresh insights into the data. The prototypes are non-incremental steps towards a final concept, iteratively informed by feedback from our domain experts and other potential future operators. 7
The converging of multiple heterogeneous data sources containing millions of nodes and edges for a graph-based research infrastructure that enables historical social network analysis creates a plethora of multidisciplinary challenges: • Dificulties associated with the merging of heterogeneous data sources • Performance of a system regarding the given scope and further scaling of the data • Creation of domain-customized interfaces, which are open and flexible with regard to unforeseen research questions • Integration of domain knowledge into the process • Visualization of millions of data points to provide explorable access points in addition to search interfaces
We address these challenges by focusing on tight, interdisciplinary collaboration and constant evaluation during the whole research and development process. Our approach brings together historical network specialists, data visualization researchers, data scientists, and experts on the evaluation of information infrastructure, an important example being the initial co-design workshop with additional external HNA practitioners and other domain experts. Building on the contextual data gathered during the co-design workshop, we will continue to follow a human-centered approach towards data modeling and visualization design.
In our next step, we aim to take a closer look at the individual processes behind historical network research in one-on-one interviews with domain experts concerning their approaches to HNA research. Our plans for the future also include the merging of multiple visualization concepts into one prototype, which will join global overviews of our data with local views of specific individual networks inside it. Furthermore, we will make use of our data and our interface to provide exemplary use cases on a variety of historical topics in collaboration with our HNA experts.
Finally, we are experimenting with linked data as an alternative to Neo4j. Here, the source data would be modeled in the form of subject–predicate– object expressions and stored in GraphDB.31 This approach would simplify the integration of Linked Open Data datasets (Wikidata, DBpedia,32 GeoNames,33 etc.), and would provide more sophisticated inference possibilities. In preliminary comparisons of the two approaches, GraphDB also shows better performance, but employing it would mean that the source data must be remodeled in order to display relation features such as relation type, relation source type, and temporal validity.
In this paper, we have described the process of examining the potential of remodeling and merging (bibliographic) big data from cultural heritage institutions into one single gathering point optimized for the use in historical network analysis. It is our hope that by providing insights into emerging challenges and outlining possible solutions, we can encourage additional research and scholarly exchange in and with similar HNA-related projects.
We would like to thank the participants of our co-design workshop and our project partners Heiner Fangerau, Katrin Getschmann, Thorsten Halling, Hans-Jörg Lieder, Gerhard Müller, Clemens Neudecker, David Zellhöfer, and Josefine Zinck. This research is part of the research project SoNAR (IDH) and is funded by the DFG – German Research Foundation (project no. 414792379).
The Discovery of Grounded Theory. Zimmerman, J., Forlizzi, J., and Evenson, S. (2007). CHI ’07: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. In Proceedings of the SIGCHI conference on Human factors in computing systems, pages 493–502. DOI: 10.1145/1240624.1240704.