Data profiling is usually performed in the very first step of a data analysis pipeline. The main goal of data profiling is to present a comprehensive overview of the data contents and its properties to the user, including any potential relationships among data attributes. In knowledge graphs, the interplay among diferent graph entities, properties, and more complex patterns that emerge in the graph is crucial to understanding its content. In this demo paper, we introduce ProGGD, a system that employs Graph Generating Dependencies to showcase information about the graph's content.
CEUR ceur-ws.org
Data profiling refers to the task of providing a comprehensive overview of the data contents
and its properties to the user. To achieve this, data dependencies are used to depict potential
correlations among the attributes and to express data quality rules. Such data dependencies
have been extensively studied in the context of relational data [
Many methods for profiling knowledge graphs have been proposed. However, few systems employ graph data dependencies to represent information about the data. The primary advantage of using graph data dependencies for profiling knowledge graphs lies in their capacity to convey detailed information about the knowledge graph to the user, and in turn, ofer insights into the quality of the data.
Graph Generating Dependencies (GGDs) [
Athlete
a debutTeam d
s SportsTeam
Athlete a nationality n
c Country s SportsTeam ciyty
ProGGD Validation of GGDs Graph Pattern Queries
Attribute Statistics
GGDMiner Attribute Similarity and
Similarity Indexes Frequent Subgraphs Discovery of GGDs
W ESR ITBEN T E A R IP FAC
E
Within ProGGD, given a knowledge graph, we identify GGDs from the data based on the frequency of a graph pattern’s appearance and the prevalence of similar or correlated attributes of nodes/edges in the graph. This discovery algorithm and the high expressivity of GGDs enables us to display information both at the schema level (the GGD itself) and at an instance level (data examples of each GGD). In ProGGD, our goal is to not only display interesting information about the data to the user but also make it easy to understand the discovered GGDs so that the user can also use it in their downstream tasks.
A GGD is defined as [] → [, ] in which [] and [, ] are respectively source
and target graph patterns and and are respectively source and target constraints on the
attributes of the nodes/edges of the respective graph patterns according to its similarity. We
say that a GGD is validated in a graph if, for every homomorphic match of the source side,
there exists a homomorphic match of the target side; we refer to [
Given a graph , we consider interesting GGDs for data profiling i.e., GGDs with graph patterns and constraints according to the similarity of the attributes (diferential constraints) that: (1) occur frequently on , (2) validated according to a user-defined rate (called confidence in our system) and, (3) can maximize the total number of matched nodes and edges in the graph (called coverage in our system). To discover such GGDs from , we use our GGDMiner discovery algorithm. According to these conditions, the main parameters of GGDMiner that need to be set by the user through ProGGD are frequency, confidence, and the maximum size of the result set of GGDs. Additional parameters such as the minimum threshold value for diferential constraints and the maximum number of edges for graph patterns can optimize the mining process. ProGGD ofer default values for these parameters to guide the users that are not familiar with the background algorithm. Due to lack of space, more information about the GGDMiner and its parameters are available in the source code repository1.
In a summary, the GGDMiner has the following steps. First, we identify the attributes of
each node/edge label that are interesting to consider for the diferential constraint discovery
and construct auxiliary data structures based on the similarity of the attributes. Next, we mine
graph patterns and constraints that occur frequently in , each combination of graph pattern
and set of constraints is considered a candidate for the source or target of a GGD. In this step,
we use algorithms and techniques for frequent subgraph mining and discovery of association
rules from the literature [
If a GGD is not validated (confidence is less than 1), it could indicate an error in the data.
ProGGD allows the user to visualize examples of validated and violated graph pattern matches
which might help the user to understand more about the content of the knowledge graph and
why such GGD was mined. ProGGD also includes functionalities to give overall information
about the graph such statistics about the attributes and graph pattern query results. We use
the Spark framework as the primary backend of ProGGD, to retrieve information about the
attributes and also query graph patterns by using the G-Core language [
Because of the rich expressiveness of GGDs, ProGGD includes many functionalities. For the user’s ease of navigation, we divide such functionalities into 4 main panels:(1) Metadata information and initialization, (2) Attribute information, (3) Topological Information and, (4) GGDs.
Metadata Information and Initialization - Given a knowledge graph and its schema information, in ProGGD we can visualize the schema and verify the attributes of each one of 1https://github.com/laricsh/ggdminer
Dataset Cordis2 GDelt3 DBLP4 the types of nodes and edges. We also display initial information about the attributes in each node/edge label to assist the user in setting the parameters for the GGD discovery algorithm. Next, we run GGDMiner to discover GGDs; this process will populate the information displayed in the other panels of the system.
Attribute Information - The GGD discovery algorithm can discover correlated attributes between the diferent nodes and edges of the knowledge graph based on its similarity. In this panel, we showcase possibly correlated attributes according to the semantic similarity of its property names, results of similarity join and clustering for string attributes, and the distribution of the values for numerical attributes.
Topological Information - In this panel, we display topological information according to the
frequent subgraph patterns mined from the knowledge graph. In this panel, user can verify
which graph patterns frequently appear and which nodes/edges are matched to each one of
these graph patterns. We use an Answer Graph [
GGDs - In this panel, we visualize GGDs discovered from the knowledge graph along with the total number of source matches, target matches and the rate of validated sources (see screenshot in Figure 2). We also show the validated and violated matches of each GGD to help the user understand the semantics of a GGD and also which subgraphs appear correlated to each other in the knowledge graph.
For the demonstration, we use open-source graph datasets and subsets of these datasets in the context of citations networks to showcase the ProGGD functionalities. Table 1 shows details about the datasets we plan to use. During demonstration, participants can select a dataset/subset, load it into the system and explore ProGGD functionalities. We showcase ProGGD in two main scenarios. In the first scenario, we focus on the topology of the knowledge graph by exhibiting mined graph patterns and their correlations using a knowledge graph with few data attributes. The second scenario, on the other hand, demonstrates the ProGGD capabilities in terms of highlighting correlations and similarities between attributes. These two scenarios illustrate how ProGGD is useful in understanding the content of the knowledge graph, even when they difer significantly in terms of topology and the number of data attributes. More details and the source code for ProGGD are available in https://github.com/laricsh/proggd. 2Graph built from Horizon 2020 project information accessed on https://data.europa.eu/data/datasets/ cordish2020projects?locale=en 3https://github.com/smartdatalake/datasets/tree/master/gdelt 4https://www.aminer.org/citation In this paper, we introduced the ProGGD system, which employs GGDs for profiling knowledge graphs. The high expressivity of GGDs enables the representation of complex information about both the topology and properties that may be associated with each other in real-world knowledge graphs. ProGGD also ofers the ability to inspect the matched nodes and edges of each GGD, thereby assisting users in understanding the content of their knowledge graph beyond mere metadata information. As part of our future work, we plan to improve ProGGD’s scalability and include functionalities that utilize GGDs in other tasks, such as data cleaning and data integration.
This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreements No. 825041 and No. 101058573.