PTSLGA: A Provenance Tracking System for Linked Data Generating Application

Published Online March 2015 in MECS

The underlying rule of the Linking Open Data (LOD) project is to provide useful and related information on the web. The data published using a Linked Data approach are open in nature, represented by RDF. Altogether, these data based on Sir Tim Berners-Lee's Linked Data principles [1]. Users can publish open data, for which, there is no guaranty of quality and accuracy of the information. Even though, publishers have been publishing their data. It contributes to an enormous growth of data on the data hub. Information may arrive from diverse sources and users can publish any kind of data without any constraint. Sometimes the links for data items found to be out-dated [2]. Due to open nature of the published data, there are various ways of revealing trust and reliability of data on the web [3]. Therefore, it brings challenging problems for the consumers to get the desired data. Subsequently all, consumer applications need information to appraise the quality of data on the web. In such instances, the publishers should provide trustworthiness and validity information at each entity and dataset level.

In this article, we present how the provenance of RDF resources is tracked and stored at the time of generating RDF resources. We also present how the provenance information is disseminated on the web. The structure of this paper is as follows. Section 2 describes some related work on this field. In Section 3, we briefly describe provenance representation and uses of widely used provenance models in the domain of Semantic Web. Section 4 explains the basic aspects of representing provenance along with conceptual terms and describes the provenance capturing architecture. In Section 5 we present the way of disseminating, consuming provenance on the web, Section 6 presents the experimental evaluation, and Section 7 concludes the work.

• II.    Related work

Linked Data generating applications need data and workflow provenance to be tracked during creation of the RDF resources in order to provide a quality provenance. The provenance needs to be managed, in the same way, as the RDF resources are available on the web. However, most of the current practices support provenance generation only at the dataset level using VoID vocabulary (Vocabulary of Interlinked Datasets) [4]. An extension to the VoID vocabulary, VoIDP, which captures the information regarding workflows and activities involved in creating an RDF dataset, is discussed in [5, 6]. A suitable provenance model that captures information about web-based and the creation of data has been elaborately discussed in [7]. One can produce provenance information at both dataset and resource level through this model. The “Named Graph” concept has also been applied to deal with the provenance information concerning the links between the data items from various sources [8].

Public Key Infrastructure (PKI) principles [9] have been used to express the trustworthiness of the dataset by using private and public keys. They suggested the use of third party trust center or Certification Authority (CA) which issues digital certificates to verify each linked datasets. These certificates have been employed to identify the validity and the quality of the datasets. A metadata component [10] for the Linked Data publishing tools such as Triplify, Pubby and D2R Server is useful for generating metadata of a huge number of RDF dataset. The component relies on the metadata provided by the publisher and allows automatic generation and publication of the provenance metadata at the dataset level. The method of automatic discovery of high-level provenance using semantic similarity has been illustrated in [11]. The methodology relies on the clustering algorithms and the semantic similarity. This approach provides provenance at the document level. In [12] authors have presented an approach on automatic generation of the metadata based on VoID vocabulary. The result produced is dataset information such as statistical data and information about linked datasets. In [13] authors have used the concept of tracking provenance information through use of Version Control System (VCS) such as Github. VCS has been mainly used in controlling and tracking source code to facilitate teamwork. Each collaborator performs some task on data or files system and commits their changes and contribution. They track such commits and perform mapping with the W3C PROV data model along with other metadata. They have provided RESTful web service to offer the provenance of such workflows. Users only need to provide URL (Github URL) that point to a Github repository. This process provides provenance at the document and file level.

A framework has been explained for converting cultural heritage data into RDF [14]. Legacy data were initially stored in spreadsheet files. The tool XLWrap has been used to translate spreadsheet data into arbitrary RDF graphs through mapping information. The converted RDF data are made available on the web accessible via SPARQL end-points. It uses two ways to disseminate the provenance or meta-data of the dataset: using VoID description of the dataset where it is published in a URL [22]. Another way is by describing used terminologies in the form of published vocabularies. For example, in [14] authors used the “?Hvor” vocabulary to represent the address information of the buildings in the Yellow List. Converting raw government data into Linked Data based on the LGD (Linked Government Data) Publishing Pipeline has been presented in [15]. The raw data are available on various formats such as JSON, CSV and TSV. They used “CKAN Extension for Google Refine” for sharing data while tracking provenance of the RDF data. This project extracts the workflow operations that are associated with the RDF data, which is purely based on process-oriented provenance. A system called BibBase [16] transforms bibliographic data stored in BibTeX files into Linked Data. RDF data are stored into a triple store which is available to be queried using SPARQL. Data in the BibBase comes from the various sources and BibTeX files. The provenance information, for each entity, is recorded by capturing the source of each entity and each link that are encoded with the entity. This method is based on data-oriented provenance where they track only the source of the data items.

• III.    Provenance Representation

The Provenance representation is a model, which assists users to express provenance of their data. Various provenance models have been evolved such as Open Provenance Model (OPM), Provenance Vocabulary, W3C Provenance Model (PROV), and VoID. The VoID is widely used in the context of Linked Data. It is a vocabulary for describing RDF dataset. VoID only deals with describing metadata of the RDF dataset. VoID has been applied to describe General Metadata (following the terms from Dublin Core), Access Metadata (information about how to locate and access the RDF data), Structural Metadata (Internal schema and technical features of the dataset), and the Description of the Linked set (a set of RDF Links). General Metadata follows the terms from Dublin Core Metadata Element set. Such as, title and description (dcterms:title, dcterms:description), licensing information (dcterms:license), creator and publisher of the dataset (dcterms:creator, dcterms:publisher). Access metadata gives information about how the RDF resources are accessed over the web. For example, SPARQL endpoints, RDF data dumps, Root resources, URI lookup endpoints, and Open Search description documents are some of the ways to access RDF dataset and RDF resources. Structural metadata provides internal structure and technical features of the dataset. For instance, the information about vocabularies used, the total number of RDF statements, entities are some of the technical features of the dataset. Oftentimes RDF resources have links to other resources from outside datasets. Such link sets are described using void:Linkset and void:target properties. Hence, VoID plays a pivotal role to distribute provenance only at the dataset level. A separate provenance model is required to represent provenance for each RDF resources in Linked Data. PROV provides core data model for representing provenance using concepts such as Entities, Activities, Agents, Roles, Time, Usage and Generations. These main concepts enlighten about how an entity came into existence, what processes and activities are involved in generating the entity, who was involved in performing the activities, what other data items were used and at what time. These key concepts take part in different aspects while performing various

Fig. 1. Describing Provenance Information using Actor, Activity and other Entities in PROV.

• IV.    Basic Aspects of Representing Provenance

• A.    RDF Statement

The abstract structure consisting of Subject (S), Predicate (P) and Object (O) is called an RDF triple. Each such triples state that the Subject and Object are in some kind of relationship joined by the Predicate (P). Such statement in RDF is called an RDF statement. Let T denotes the RDF Triple, then

T = 5 U P и о                      (1)

activities. Here, the main activity is data conversion, which converts source data items into target data items. Users who are involved in this conversion process are agents, which in turn are associated with a particular agent, an organization. So, using all these concepts publisher can model provenance of a data item. Fig. 1 shows the detail concept about how a data item is related to an activity, agent and other data items. To represent provenance of RDF resources we use PROV data model. VoID vocabulary is used to describe the provenance of the RDF dataset.

In RDF, the subject is always the resource that is being described. A resource can be of anything, a place, a person, and a book such that subject and predicate are always identified by URI whereas the object, which can be either a resource (identified by URI) or a literal value. In this work, we denote a subject by R.

• B.    RDF Dataset

RDF dataset is a collection of RDF triples or statements. It is also called a directed or labeled graph where subjects and objects are nodes and the predicate represents the arc. Let {T 1 , T ₂,... ,T _n } be a set of RDF triples. The RDF dataset D is defined as:

D = { T„ T ₂,..., T, }                             (2)

• C.    Provenance of Data Item

In generating data-items of any kind, the data item is associated with many things such as agents, activities, processes, and other used data-items. Hence, provenance of a data item, in the perspective of the agent, entity and process oriented provenance, is a collection of agents, activities, processes, and source data items. We represent the provenance for any data-item by P such that,

P = [ A , V , F , 5 ]                              (3)

Where,

A = { A 1 , A 2 ... A n } is the set of all agents,

V = { V 1 , V 2 ... V n } is the set of all activities,

F = { F 1 , F 2 … F n } is set of functions or processes that generates the data-item or relates a data-item to another data-item.

S = [ S₁ , S₂ … S_n ] is the collection of source of the data item, other used data items and the brief description about the origin of the data.

• D.    Provenance of RDF Resource

Provenance of RDF resource may be defined as tracking provenance of R where R is the resource or subject in an RDF triple. Provenance of R is defined as the combination of its agent, process, entity oriented provenance and the provenance generated by the previous data storage system. Such that,

Pr = P • Pl                              (4)

Where,

P is the agent, process, and entity oriented provenance.

P L is the provenance information recorded by the previous system for legacy data.

• E.    Provenance of RDF Dataset

The provenance of an RDF dataset is a collection of general metadata, access metadata, structural metadata and the description of link set as well as the combination of its agent, process, and entity-oriented provenance, which is described by VoID. Such that,

P d = [ G, а , ^ , Я ] * P                       (5)

Where,

G is general metadata,

α is access metadata, ψ is structural metadata λ is the description of the link set, and P is the agent, process, and entity oriented provenance.

The comprehensive architecture of the framework has been shown in Fig. 2. Based on the above concepts, for each data item, we capture the provenance information such as the source of each statement, associated activities, agents, the date-time information, the actors and the information about processes that were involved in the creation process. The source of the data provides information, which corresponds to the creational history of the data. The source of the data may not be available to the process and therefore the agent needs to enter it explicitly. The agent enters basic information such as the source of the input data, agent's address and the description about the input data and licensing information. Sometimes the provenance is also captured and recorded in the legacy data system. For example, in case of MARC 21 record, the Field 561 defines it as “Ownership and Custodial History” [20]. The provision for storing such information has been made and is combined with the agent, process, and entity oriented provenance. The agent is associated with the data creation and modification process. It is automatically created and stored into the store.

Fig. 2. Provenance Capturing Architecture.

Whatever data are being created they are associated with only one agent. The data conversion activity is the process, which creates data item and assigns relevant provenance information to it. Hence, the source data, date-time parameters, the agent, activities, and other entities are required to capture the provenance information. In the end, the conversion process collects all relevant metadata of the dataset and creates VoID file separately. By this, the Linked Data generation application becomes conscious towards provenance, making provenance tracking as a mandatory step for them. As shown in Fig. 2, the legacy data along with source information, agent's information as well as licensing information is entered by the user. In the next step, data analyzer analyses the data to be processed. This will verify whether the supplied data can be processed by the data conversion activity. The data conversion activity takes each data item, converts into RDF and Linked Data. As a whole, the provenance information is recorded by the provenance capturing method. A separate method is added which performs the job of provenance capturing. During the conversion process, for each data item, the corresponding provenance item is created and it’s certain attributes such as date-time, activities, and agents are assigned to the provenance data item. Any source data, which are linked to outside sources, such as rdfs:seeAslo, owl:sameAs links are treated as used data items. The provenance store communicates inwardly with the RDF store to retrieve external data sources that have been linked with the data. Each time the legacy data item is encountered, it is converted into RDF resource, integrated with other data sources from the web such as DBpedia, VIAF and then it is stored into the RDF store. Simultaneously, the provenance generator captures the provenance of the data item and stores into provenance store.

• V. Provenance Publication and ConsumPTION

Publishing provenance is a vital job for the data publishers. Making the availability of the provenance information on the web is always a concern. In some measure, publishers should provide references or access information of the location of provenance for each data item. Provenance is valuable information that one must provide its access in order to ensure that data is trusted. Several choices have been made regarding provenance publication. Oftentimes VoID is used to express metadata of Linked Data. Using VoID, we can publish the provenance of the RDF dataset as a separate document on the web and then it is linked from the RDF document using void:inDataset property as discussed in [17]. The Provenance Access and Query (PAQ) working draft [18] have elaborated a number of possible ways for accessing provenance for individual data items. It is mentioned that the provenance information should be accessible, in the same way, as the resources are accessible on the web, by dereferencing the HTTP URI. It means that the provenance information is also a resource represented and described by RDF having dereference-able HTTP URI for each resource. The following are the two different ways for accessing the provenance:

• A.    Indirect Access

When provenance is not accessible using provenance-URI, or they are not accessible as the web resource, a Query Service can be used to serve the provenance information. Such query service, such as SPARQL service endpoint, processes the SPARQL queries submitted by the consumers. In such cases, the publishers should mention the provenance service URI in the RDF resource and HTTP response header field for the resource represented by RDF and HTML respectively.

• B.    Direct Access

Another way to access provenance information of the RDF resources is by dereferencing provenance-URI. Provenance-URI points to the actual provenance record generated and stored by the data publisher. Dereferencing this URI discovers the provenance information associated with the original RDF resource. Provenance URI has been embedded in the original RDF resource using “prov:hasProvenance” property for the resource, represented as RDF. It can also be associated with a resource by adding element followed by

in the HTTP response header field for the resource represented as HTML.

In the proposed work, we follow the direct approach. Provenance information for each resource has been defined using RDF following the PROV data model. An RDF resource represented by the URI has the resource description. In the resource, reference to the provenance URI has been added using property followed by the provenance URI . Dereferencing the provenance URI results in the provenance description in RDF format considering the “Accept-Encoding” of the HTTP header field as “RDF/XML”. The provenance description tells that, the resource

prov:startedAtTime “Sat Jul 05 23:29:45 GMT+05:30 2014”, prov:wasGeneratedBy the activity , prov:wasDerivedFrom the dataset , it prov:used the resource "" for generating the links and was prov:endedAtTime “Sat Jul 05 23:29:49 GMT+05:30 2014”. This resource prov:hadPrimarySource < > and has DC:provenance as “Boston, Mass.” from the previous data storage system which was published by “American Society of Landscape Architects”. The activity whose title is “Legacy Data to Linked Data Converter” prov:wasAssociatedWith , created by “University of Kalyani”. The agent is type of “prov#Person" and was prov:actedOnBehalfOf another agent of type “prov#Organization". In this way, provenance has been generated for each resources.

Table 1. Evaluation Results (Legacy Record to RDF with Provenance)

Dataset	No. of Legacy Record	Time to convert Legacy record to RDF without Provenance (in Seconds)	Time to convert Legacy record to RDF with Provenance (in Seconds)	Number of RDF Resource (in %)
Dataset 1	100000	28s	52s	96%
Dataset 2	100000	24s	50s	97%
Dataset 3	100000	21s	45s	87%
Dataset 4	100000	16s	39s	100%
Dataset 5	100000	18s	40s	94%
Dataset 6	100000	23s	48s	87%

• VI.    Experimental Evaluation

Harvard Library Bibliographic Dataset has been used as legacy dataset, provided by Harvard Library [21] for the sake of experiment. Library aims at providing open metadata in the library domain to support learning and research for the students as well the researchers. The datasets consist of bibliographic records in MARC 21 format, exported from Harvard's Library for public use. The datasets contain more than 12 million bibliographic records of different categories such as books, journals, electronic resources, audios, videos and other materials. The proposed work is an extension of the previous work [19]. The method has been improved and it takes less time to process legacy resources than previous work. Six different bibliographic datasets have been processed for the sake of experiment. Each dataset is of different size, from which a varied number of RDF resources and provenance information have been generated. As shown in the Evaluation Table 1, to convert first 100000 legacy records without having provenance and links to outside sources it took less than 30 seconds. However, it took around 50 seconds with provenance records but without having links.

Table 2. Evaluation Results (Legacy Record to Linked Data with Provenance)

Dataset	No. of Legacy Record	Time to convert Legacy record to Linked Data without Provenance (in Seconds)	Time to convert Legacy record to Linked Data with Provenance (in Seconds)	Number of RDF Resource (in %)
Dataset 1	100000	31243s	31904s	96%
Dataset 2	100000	27209s	29138s	97%
Dataset 3	100000	28971s	31278s	87%
Dataset 4	100000	30929s	39614s	100%
Dataset 5	100000	32138s	33860s	94%
Dataset 6	100000	19260s	21156s	87%

The experiment has been performed on a 64-bit 2 GHz Intel Core i7 processor having 4GB of RAM running on Mac OS X 10.8.3. Jena 2.10 framework has been chosen to process RDF data along with Jena TDB to store RDF and provenance resources.

Fig. 3. Legacy Record to RDF and Linked Data conversion for first 100000 records.

• VII.    Conclusion

In this article, we have shown how the provenance information of the RDF resources can be generated during the conversion process. Since many Linked Data generating applications are not aware of provenance information of RDF resources. During the conversion of legacy data to Linked Data the provenance needs to be captured and recorded instantly. Many consumer applications need a technique to trace metadata or provenance of the data items on the web. For this we have followed the direct approach, mentioned in

Provenance Access and Query (PAQ) working draft [18], to disseminate the provenance information. We have shown how an RDF resource can be linked to its provenance-URI. Provenance-URI is dereference-able on the web and is accessible by both HTTP and RDF. We believe that by tracking provenance of RDF resources during the conversion process will help publishers in generating provenance while reducing the time and efforts. The future work includes the conversion of the legacy data from other data formats such as CSV keeping their provenance, versioning and change information.