In this vision paper we present a direction for visualizing the implicit schema of a loosely-defined RDF graph in diferent zoom levels. RDF datasets do not necessarily adhere to strict structural guidelines, even though a schema in the traditional sense (e.g., an underlying ontology or simple rdf:type declarations) usually exists. Instead, there can be many diferent ways that instances of a particular type can appear in the same dataset. For example, two instances of the same type can have diferent properties, thus making their type equality semantically ambiguous. Visualizing this schema can lead to either non-intuitive depictions of the implicit types, or over-generalized visualizations that lack informative detail. In this paper, we propose our vision of an approach that tackles this issue by embedding the implicit schema into a vector space that captures generalizations and specializations of the implied types and relationships in an RDF graph.
The Linked Open Data cloud ofers increasingly large amounts of RDF1 datasets, covering a wide range of domains and topics that include life sciences, demographics, government data and general encyclopedic knowledge. In contrast with the more traditional tabular data management paradigm that usually enforces structural restrictions, RDF represents relationships between concepts in the form of a graph that is often loosely defined; that is, arbitrarily diverse types of concepts and relationships can co-exist in the same RDF dataset without loss of consistency. Unfortunately, this heterogeneity in the schema, while diverse and dynamic, can lead to problems not only in storage, indexing and querying, but to schema comprehensibility and user-friendliness as well, as users cannot intuitively explore the dataset schema.
As an example, consider Figure 1(a), where two instances of rdf:type foaf:Person are depicted. While Alice and Claire are both instances of the foaf:Person class, they do not share the same properties, except for the rather generic rdf:type property. A question worth asking is: how do we visualize the foaf:Person class? In a relational context, this would amount to two diferent table definitions for the implied types of Alice and Claire, even though in the RDF dataset they are declared to belong in the same class. If we were to visualize just one class (i.e., foaf:Person), this would hide the fact that the structure of the instances of this class is vastly diferent 2. Thus, it may be more intuitive to have two classes in the visualization, which would better capture the sets of relationships, or properties, coming out of each implied type.
Similarly, consider the case depicted in Figure 1(b). In this case, Alice and Claire again belong in the same class, but this 1https://www.w3.org/RDF/ 2In this example, it appears that Alice is represented by her professional attributes, while Claire is not. (a) Alice and Claire have entirely diferent property sets
(b) Alice and Claire have overlapping property sets time there is an overlap in their properties. More specifically, Alice and Claire both share the worksFor property, but Alice is also described as a supervisor of Beth. This can be interpreted as follows: the two instances belong in the same class at some height of the class’s hierarchy, before they branch out to diferent specializations. For instance, Alice appears to be a supervisor or a manager, while Claire appears to be a more generic type of employee. Therefore, asking how to visualize the foaf:Person class is again worthwhile. Visualizing the schema with just one class is misleading, as there may be more hidden specializations. While this latter point can be addressed with merging/splitting the class node in diferent zoom levels, the examples aim to show that a somewhat trivial RDF dataset can hide more detailed schema information that what is explicitly declared.
Recent works in the state of the art in RDF data
management have shown that exploitation of an implicit schema of the
data can be beneficial in several diferent directions, such as
storage, querying and schema exploration [
In this sense, CSs can be thought of as implied types of nodes. However, they do not necessarily capture compact and comprehensible structural information. For example, some commonly used and heavily cited RDF datasets include CSs in the order of 103 − 104 3. Thus, while using CSs to visualize the RDF schema of such a dataset is more informative than just showing the structural definitions that explicitly exist in the dataset (e.g., all objects of rdf:type), such a visualization would neither be intuitive, nor comprehensible, as it would lead to huge graphs of implied types and relationships between them.
In order to both use CSs and CS relationships for schema summarization and consequent visualization, and address the issue of large amounts of implied types, in this paper we envision an approach that embeds CSs in a continuous vector space and uses the distances between vectors in order to provide a means for merging similar CSs (when zooming out) and splitting merged CSs to their diferent specializations (when zooming in). Our approach aims to take into account latent semantic similarities between implied types (CSs). For this reason, we propose to use a continuous vector space model where CSs can be embedded.
There are two main directions that can be followed here; first,
we can use a pretrained model of word vectors, such as the Google
News Corpus model, which contains a vocabulary of 3 million
unique words, trained on an input corpus of 100 billion words.
Second, we can build a vector space directly by embedding CSs,
by following a graph embedding approach as is done in recent
works[
This paper is structured as follows. Section 2 describes the main concepts of the approach, section 3 describes diferent embedding methods and section 4 concludes the vision paper. 2
In this work, we consider characteristic sets CSs as nodes in a graph, with an edge between two CSs denoting that there is a property that connects two subjects that belong in two diferent CSs. Formally, given an input RDF dataset D, a CS graph GD is an ordered pair GD = (VD , ED ), where VD is the set of all CSs in D, and ED ⊆ VD ×VD is the set of all properties that connect pairs of CSs. The key idea of our approach is to extract the CSs of an RDF dataset, derive an implied schema and visualize the resulting implied schema. In this sense, we consider a CS to define an implied type, i.e., a non-explicitly defined class of instances. However, as was discussed in the introduction, this entails several problems. It does not sufice to extract all CSs and visualize them in a graph of CS nodes and CS-to-CS edges, because their heterogeneity in real-world datasets makes the approach prohibitive. It would not, for example, make sense to show 150 diferent CSs that all essentially describe the same class, but with slight variations in their property sets, especially when exploring the schema visually and at a distance (e.g., in high zoom levels); instead, it would be more meaningful to specialize/generalize the multiple CSs in diferent zoom levels. Consider the example of Figure 1(a) and 1(b). The CSs and how they are merged in diferent zoom levels can be seen in Figure 2.
Of course, it would not be reasonable to merge all CSs in one node at the top zoom level, because many of them describe diferent types, either implicit or explicit. For this reason, we need to apply a similarity metric that can be parameterized with the zoom level. Based on this metric, we can merge together only the CSs that are close to each other at each level of visualization. This procedure can be interpreted as a form of hierarchical clustering.
While graph visualization, hierarchical visualization and
visual LOD exploration have been thoroughly studied recently
3DBPedia has more than 10,000 CSs, while Geonames and WordNet have ~1000
[
In this direction, we propose to use embeddings in order to
position the CSs in a continuous space, and then rely on their
geometric distances in order to attribute similarity or diference.
The advantage of the use of the embedded space is that it can
provide latent semantic similarity characteristics to the CSs and
thus associate even seemingly unrelated CSs. For example,
consider two CSs, c1 = {worksFor, specializesIn, hasSalary} and
c2 = {hasPro f ession, hasIncome, isEmployedAt } that share no
common properties. As can be seen though, there is a latent
semantic association between the two CSs, as we can assume that
they both describe people that are professionals. Thus, it would
make sense to merge c1 and c2 together when visually exploring
the schema at a distance. This type of similarity is easily
captured with the help of embeddings [
The two main directions that will be pursued are (i) mapping CS nodes to an existing vector space, and (ii) embedding CS nodes directly in their own respective vector space. In what follows, these two cases will be discussed. 3.1
An embedding of an item (a word, a concept, a graph node etc.) in a continuous vector space is essentially the positioning of the item in a multidimensional space of arbitrary dimensions, where its locality is afected by the similarity of the neighboring items. In other words, the item is assigned a vector that represents its position in the space, so that all surrounding items that are also represented by vectors are semantically similar.
There are many diferent ways to define similarity. One of the
most adapted and promising ways is to define the similarity of
two items based on their context. In the case of text, where each
item is a word or a phrase, the context is defined as the word’s
or phrase’s surrounding words. Thus, two words that commonly
coexist in the same sentences will generally result in neighboring
vectors. The word2vec model [
In our approach, the key idea is to directly map CSs to vectors in a pre-trained linguistic vector model that already contains latent semantics for the whole English language. In order to do so, we need to follow the following steps: (1) Assign a set of descriptive words to each CS (2) Aggregate the vectors of the assigned words to derive a single vector (3) Map the CS to the aggregate vector
For step (1), we can use the property set and the assigned rdf:type classes (if any) of the CS to come up with descriptive words. For example, for c1 in Figure 1(b), we can use the words person, works and supervises. Then, for step (2), we first look up the existing word2vec model to find the vectors for person, works and supervises. Then, we can find a single vector representation of the three words by performing an operation such as mean or sum of their respective vectors. Finally, this aggregate vector will be the embedding of the CS. We can then define appropriate distance thresholds to find similar or dissimilar CSs.
3.1.1 Google News Corpus Example. To highlight the advantage of using such an NLP model to associate CSs in contrast with strictly relying on set operations on the CSs’ property sets, let us study an example. As a pretrained NLP model, we will use the vectors created by word2vec for the Google News Corpus. This model contains vectors for a vocabulary of 3 million English words in a 300-dimensional continuous vector space, and was trained over more than 100 billion English words from news articles. Thus, it provides ready-to-use English language vectors with general, context-independent associated semantics.
For our example, let us consider an RDF dataset, whose resulting CS graph contains the following CSs: (1) c1 = {worksFor, specializesIn, hasSalary} (2) c2 = {hasPro f ession, hasIncome, isEmployedAt } (3) c3 = {worksFor, name, aдe } (4) c4 = {hasSubOrдanization, reдisteredAt, hasCEO } (5) c5 = {hasMinister, belonдsInGovernment }
By inspecting the CSs, a human analyst can conclude that
c1, c2 and c3 describe people, while c4 and c5 are more related
to organizations (companies and ministries). Let us also assume
that the first three CSs are also associated with the foaf:Person
class in the RDF dataset, i.e., the subject nodes of the RDF graph
that belong in these CSs are of the type foaf:Person, while c4 and
c5 are of types Company and Ministry respectively. In recent
approaches such as [
After some preprocessing for step 1 that was described before, we can derive descriptive words for the five CSs. Let us assume that these are c1 : {works, specializes, salary, person}, c2 : {profession, income, employed, person}, c3 : {works, name, age, person}, c4 : {suborganization, registered, CEO, company} and c5 : {minister, government, ministry}. Feeding these into the pretrained Google News word2vec model and computing the mean vector for each of the five sets, we get the pair-wise similarity matrix shown in Table 1.
As can be seen, c1 and c2 are more closely related with each other than with c3, even though they do not share any common properties. Interestingly, c1 and c2 are more similar than c1 and c3 even though the latter actually share a common property. Furthermore, c1, c2c3 are closer with each other than with c4, c5, while c4 and c5 are more similar to each other than with the rest of the CSs. This intuitive knowledge that can be used to cluster together similar CSs based on their latent semantics would not have been captured using traditional methods that rely on property set operations. Using this approach, a potential visualization of this schema in diferent zoom levels can be seen in Figure 3.
Concluding this example, it should be mentioned that in order to serve more context-dependent data, such as specialized scientific data, we can define and configure the training corpus accordingly. For example, for life sciences datasets, it might not make much sense to use the Google News Corpus. Instead, it could be more useful to build a corpus by combining life science text sources, such as UMLS, Drugbank and related Wikipedia subsets, than rely on a generic and multi-purpose vector model.
A recently growing body of work focuses on so called graph
embeddings. The objective of these approaches is to embed a
graph of nodes and edges into a continuous vector space, while
preserving the inherent structure of the graph. Most of the related
work focuses on preserving this structure for the nodes of the
input graph [
However, these approaches operate on the instance level and are not applied to the schema of the graph, as is our case. As a consequence, they train on the whole dataset’s instance, and thus they rely on much larger corpora to build their vector models, in contrast with our case where the implied schema is usually orders of magnitude smaller than the size of the dataset. Thus, the encoded information and relationships will be supported by an abundance of data and hence will potentially be more accurate.
On the other hand, training on just the CS graph can be a computational advantage of using graph embedding techniques, in that the training of the model and construction of the vector space will be based only on the input CS graph, where in our case it is expected to be small and proportional to the amount of CSs in a dataset4. Thus, it would not be computationally costly to train the vectors or perform neighbourhood queries for a given CS. A problem with this approach, however, is that the output model is dependent on the input and will be very limited in discovering latent semantic similarities between seemingly unrelated CSs.
In this vision paper, we have presented a potential approach for visualizing a loosely-defined RDF schema by combining characteristic sets with embeddings. To this end, we have proposed two directions, (i) by using NLP embedding models such as word2vec, and (ii) by using graph embedding techniques. We focus our vision on taking advantage of the latent semantics that can be provided by embeddings in order to provide visual generalizations and specializations of an RDF dataset’s implied types in diferent zoom levels.
So far, we have only discussed grouping and visualizing implied types based on CSs. In the case of relationships between CSs, we can perform the same steps, but this time instead of CSs we will consider groups of properties that connect two CSs in the dataset. Then, we can group/ungroup the resulting edges accordingly, based on the same methodology as above.
Another point that requires addressing is the fact that proper diferentiation of the implied types is not possible by just relying on metric distances. For example, the cosine similarity between c1 and c4, i.e., between human professionals and companies in our example is actually not very diferent than the similarity between c1 and c3. Thus, we would not be able to place humans and organizations in diferent groups by just relying on their similarity in the vector space. This would require auxiliary knowledge 4Theoretically, a CS graph with n CSs can have up to n2 edges, but this would mean that each implied type of concept is associated with all of the rest of the types, which is not realistically expected in common datasets or better vector assignment. In our case, the existence of rdf:type properties can serve as the auxiliary knowledge. However, this information will not always be available, thus this point needs proper addressing in the future.
Finally, grouping at diferent levels will be addressed by an appropriate clustering approach, which will require tuning of the required hierarchical thresholds, or the ranges of similarities serving each zoom level. For example, in preliminary experimentation, trivial methods such as linear partitioning of similarity ranges can be used in order to assign ranges of similarity for diferent zoom levels (e.g., for the lowest zoom level we group together all CSs with centroid-based similarity larger than 0.9), and loosen this value the higher we go in the zoom level. As a ifnal note, appropriate hierarchical or centroid-based clustering methods will be studied for the above purposes.
In conclusion, it is worth mentioning that the proposed
approach can be applied in more fields than schema visualization.
Schema summarization, indexing and query planning are a few
domains that can take advantage of the proposed technique. For
example, in [
Acknowledgements. We acknowledge support of this work by the project "Moving from Big Data Management to Data Science" (MIS 5002437/3) which is implemented under the Action "Reinforcement of the Research and Innovation Infrastructure", funded by the Operational Programme "Competitiveness, Entrepreneurship and Innovation" (NSRF 2014-2020) and coifnanced by Greece and the European Union (European Regional Development Fund).