The ready availability of data is leading to the increased opportunity of their re-use for new applications and for analyses. Most of these data are not necessarily in the format users want, are usually heterogeneous, and highly dynamic, and this necessitates data transformation e orts to re-purpose them. Interactive data transformation (IDT) tools are becoming easily available to lower these barriers to data transformation e orts. This paper describes a principled way to capture data lineage of interactive data transformation processes. We provide a formal model of IDT, its mapping to a provenance representation, and its implementation and validation on Google Re ne. Provision of the data transformation process sequences allows assessment of data quality and ensures portability between IDT and other data transformation platforms. The proposed model showed a high level of coverage against a set of requirements used for evaluating systems that provide provenance management solutions.
The growing availability of data on the Web and in organizations brings the opportunity to reuse existing data to feed new applications or analyses. In order to reuse existing data, users must perform data transformations to repurpose data for new requirements. Traditionally, data transformation operations have been supported by data transformation scripts organized inside ETL (ExtractTransform-Load) environments or by ad-hoc applications. Currently, users developing data transformation programs follow a typical software development cycle, taking data samples from datasets, and developing the transformation logic, testing and debugging. These approaches are problematic in emerging scenarios such as the Linked Data Web where the reduction of the barriers for producing and consuming new data brings increasing challenges in coping with heterogeneous, high-volume, and dynamic data sources. In this scenario, the process of interactive data transformation emerges as a solution to scale data transformation.
Recently platforms, such as Google Re ne3, are exploring user interface and interaction paradigms for de ning data transformations. These platforms allow users to operate over data using Graphical User Interface (GUI) elements instead of scripts for data transformation. By providing a set of pre-de ned operations and instant feedback mechanism for each iteration of the data transformation process, this model de nes a powerful approach to data transformation and curation. However, despite its practical success, the assumptions behind platforms such as Google Re ne have not been modeled nor formalized in the literature. As the need for curation increases in data abundant environments, the need to model, understand, and improve data curation and transformation processes and tools emerges as a critical problem. A foundational model of IDT that (a) brings out the underlying assumptions that IDT platforms use, (b) makes explicit how IDT relates to provenance, and (c) helps IDT platforms be comparable and thereby helping in de ning interfaces for their interoperability, would be highly useful.
The process of data transformation is usually positioned as one element in a connected pipeline which needs to be integrated with di erent processes, to form a work ow. Currently, the set of transformation operations present in IDT platforms are not materialized in a way that could allow its use in contexts outside the IDT environment. A transformation process normally involves a data object that is being transformed, and the transformation procedure itself. Provenance is the contextual metadata describing the origin or source of that data. Prospective provenance provides mechanisms to describe generic work ows which could be used to materialize (future) data transformations. Retrospective provenance captures past work ow execution and data derivation information to provide useful context for what had happened up to the present state and time.
This paper investigates the two complementary perspectives described above, and an approach for expressively capturing and persisting provenance in IDT platforms. A provenance extension for the Google Re ne platform is implemented and used to validate the proposed solution. The supporting provenance model focuses on the maximization of interoperability, using the three-layered data transformation model proposed in [?] and uses Semantic Web standards to persist provenance data.
The contributions of this work include: (a) a formal model of IDT; (b) the use of an ontology-based provenance model for mapping data transformation operations on IDT platforms; (c) the veri cation of the suitability of the proposed provenance model for capturing IDT transformations, using Google Re ne as the IDT platform; and (d) the validation of the system against the set of require
Prototypical User Actions Data Curation Application Operation selection
KBwoforokpfelorawtisons/ Selection Parameter selection (extensional semantics)
Selection Operation composition
Storage Prospective provenance
Materialization
Data
Schema-level Data Instance-level Data
Input Data Transformation Generation Cycle Data Input (Re)Definition of User Actions Transformation Execution Data
Output Inductive assessment of user intention
Output Data Transformation Program
Parameter α Operation A Parameter β
...
Parameter ω
Operation B
...
Operation Z
Retrospective provenance
Materialization Output
Data Specific Transformation
ODuatptaut trDawantsoafroskrpfmelocawitfiiocn ments, from the literature, used for evaluating provenance management provided by IDT platforms.
Interactive Data Transformations (IDT)
IDT is de ned as the application of a pre-de ned set of data transformation operations over a dataset. In IDT, after a transformation operation has been selected from the set of available operations (operation selection), users usually need the input of con guration parameters (parameter selection). In addition, users can compose di erent set of operations to create a data transformation work ow (operation composition). Operation selection, parameter selection, and operation composition are the core user actions available for interactive data transformation as depicted in gure ??.
In the IDT model, the expected outputs are a data transformation program and a transformed data output ( gure ??). The transformation program is generated through an iterative data transformation program generation cycle where the user selects and composes an initial set of operations, executes the transformation, and assesses the suitability of the output results by an inductive analysis over the materialized output. This is the point where users decide to rede ne (reselect) the set of transformations and con guration parameters. The organization of operations as GUI elements minimizes program construction time by minimizing the overhead introduced by the need to ensure programming language correctness and compilation/deployment.
The transformation generation cycle generates two types of output: a data output with a data speci c transformation work ow, and a data transformation program which is materialized as a prospective provenance descriptor. A provenance descriptor is a data structure showing the relationship between the inputs, the outputs, and the transformation operations applied in the process. While the provenance-aware data output is made available for further changes, the prospective provenance work ow is inserted into the KB of available workows, and can be later reused.
In the next section, we shall present an algebra for Interactive Data Transformation. 2.2
Despite the practical success of IDT tools, such as Google Re ne, for data transformation and curation, the underlying assumptions and the operational behaviour behind such platforms have not been explicitly brought out. Here, we present an algebra of IDT, bringing out the relationships between the inputs, the outputs, and the functions facilitating the data transformations.
gine G A provenance-based Interactive Data Transformation Engine, G, consists of a set of transformations (or activities) on a set of datasets generating outputs in the form of other datasets or events which may trigger further transformations.
G is de ned as a tuple,
G = < D; (D [ V ); I; O; ; ; > where 1. D is the non-empty set of all datasets in G, 2. D is the dataset being currently transformed, 3. V is the set of views in G (V may be empty), 4. I is a nite set of input channels (this represents the points at which user interactions start), 5. O is a nite set of output channels (this represents the points at which user interactions may end), 6. is a nite alphabet of actions (this represents the set of transformations provided by the data transformation engine), 7. is a nite set of functions that allocate alphabets to channels (this represents all user interactions), and 8. = < D O ! > is a function, where, in a modal transformation engine, (D; O) 2 (O) is the dataset that is the output on channel O when D is the dataset being currently transformed.
Transformation Event is a tuple,
PT E = < Di; Ftrans; (Do [ V ); Ttrans > where { Di is the input dataset for the transformation event, { Do is the dataset that is the result of the transformation event, { V is a view or facet that is a result of the transformation event, { Di [ Do [ V D { Ftrans is the transformation function applied to Di (applied element-wise), and { Ttrans is the time the transformation took place.
De nition 3: Run A run can be informally de ned as a function from time to dataset(s) and the transformation applied to those dataset(s). Intuitively, a run is a description of how G has evolved over time.
So, a run over G is a function, P : t ! < D; f > where t is an element in the time domain, D is the state of all datasets and views in G, and f is the function applied at time, t.
A system R over G, i.e. R(G) , is a set of all runs over G. We say that < P; t > is a point in system R if P 2 R.
P captures our notion of \prospective provenance".
De nition 4: Trace Let =< P; t >2 R(G). The trace of , denoted by, !, is the sequence of pairs < ri; ti > where ri is the i th run at time ti. The set of all traces of G, T r(G), is the set f ! j 2 R(G) g. An element of T r(G) is a trace of G. A trace captures our notion of retrospective provenance. 3 3.1
Provenance-based Data Transformation
Community e orts towards the convergence into a common provenance model led to the Open Provenance Model (OPM)4. OPM descriptions allow interoperability on the level of work ow structure. This model allows systems with di erent provenance representations to share at least a work ow-level semantics. OPM, however, is not intended to be a complete provenance model, demanding the complementary use of additional provenance models in order to enable uses of provenance which requires higher level of semantic interoperability. This work targets interoperable provenance representations of IDT and ETL work ows using a three-layered approach to represent provenance. In this representation, the bottom layer represents the basic work ow semantics and structure provided by OPM, the second layer extends the work ow structure provided by OPM
with Cogs[?], a provenance vocabulary that provides a rich type structure for describing ETL transformations in the provenance work ow, and voidp [?], a provenance extension for the void5 vocabulary, that allows data publishers to specify the provenance relationships of the elements of their datasets. The third layer consists of a domain speci c schema-level information of the source and target datasets or classes/instances pointing to speci c elements in the ETL process. In our implementation, the third layer contains the mapping of instances to source code elements. Figure ?? shows how we applied the model (and architecture) to Google Re ne, a popular IDT platform. 3.2
There are two major approaches for representing provenance information, and these representations have implications on their cost of recording. These two approaches are: (a) The (Manual) Annotation method: Metadata of the derivation history of a data are collected as annotation. Here, provenance is pre-computed and readily usable as metadata, and (b) The Inversion method: This uses the relationships between the input data, the process used to transform and to derive the output data, giving the records of this trace.
The Annotation method is coarser-grained and more suitable for slowlychanging transformation procedures. For more highly-dynamic and time-sensitive transformations, such as IDT procedures, the Inversion method is more suitable. We map the provenance data to the three-layered data transformation model provenance model as described in section ??.
After choosing the representation mechanism, the next questions to ask are: (a) what data transformation points would generate the provenance data salient to our provenance needs, and (b) what is the minimal unit of a dataset to attach provenance to. For our system, we choose an Interactive Data Transformation Event to capture these two questions and this is made up of the data object
being transformed, the transformation operation being applied to the data, the data output as a result of the transformation, and the time of the operation (as stated in section ??).
User
User Interactions User
Queries Queries
Provenance Graph Provenance Storage Layer
Data Transformation
Engine Provenance Event Capture
Layer Provenance Representation
Layer
Tt Di Do Ft
Interceptor Data Transformation
Engine
Google Refine Keys: Tt: Time of Operation Di: Input Data or Dataset Do: Output Data of Transformation Ft: Applied Transformation Operation
Tt Di Do Ft
Object-toJavaClass Introspector Event-to
RDF Statements Mapper Provenance Data in RDF Provenance Storage (A) Process Flow
(B) Provenance Event Capture Sequence Flow 1. The Provenance Event Capture Layer, which consists of the following layers and operations' sequence ( gure ??(B)): (i) The Interceptor Layer: Here, user interactions are intercepted and provenance events extracted from these interactions; (ii) Object-to-JavaClass Introspector: The inputs to this layer are the Transformation Operation chosen to transform the data. We employ the Java language re ection mechanism to elicit the full class path of the operation performed. Eliciting the full class path is useful for the following reasons: (a) It allows us to have a pointer to the binary of the programs doing the transformation, and (b) This pointer to the program binary allows the connection between the full semantics of the program and the data layer. The outputs of this layer are the full class paths of the transformation operations; (iii) Event-to-RDF Statements Mapper: it receives the provenance event and is responsible for the generation of the RDF predicates. 2. These events are then sent to the Provenance Representation Layer, which encodes the captured events into RDF using the provenance representation described in section ??. 3. These events, represented in RDF, are then sent to the Provenance Storage Layer, which stores them in its Knowledge Base (KB). 4
Google Re ne6 (GRe ne) is an exemplar Interactive Data Transformation system, and some of its mechanisms include: (i) ability to import data into GRe ne from di erent formats including tsv, csv, xls, xml, json, and google spreadsheets ; (ii) GRe ne supports faceted browsing7, such as: Text Facets, Numeric Facets, and Text Filters; (iii) Editing Cells, Columns, and Rows, using GRe ne's editing functions; and (iv) The provision of an extension framework API. 4.1
Some of the design decisions that must be made when making an application provenance-aware is to ask \what" type of provenance data to capture, \when" to collect the said provenance data, and \what" type of method to use for the capture.
As regards to \when" to collect provenance data: (a) provenance data can be collected in real-time, i.e. while the work ow application is running and the input dataset(s) are being processed and used, or (b) ex-post (after-the-fact), i.e. provenance data is gathered after a series of processing events or a sequence of activities has completed.
As regards to \what" type of method to use for collection, provenance data collection methods fall into three types: (a) Through \User annotation": A human data entry activity where users enter textual annotations, capturing, and describing the data transformation process. User-centered metadata is often incomplete and inconsistent [?]. This approach imposes a low burden on the application, but a high burden on the humans responsible for annotation; (b) An automated provenance instrumentation tool can be provided that is inserted into the work ow application to collect provenance data. This places a low burden on the user, but a higher burden on the application in terms of process cycles and/or memory, and (c) Hybrid method: This method uses an existing mechanism, such as a logging or an auditing tool, within the work ow application to collect provenance data.
We built an automated instrumentation tool to capture provenance data from user operations as they use the system ( gure ??). This approach incurs very little burden on the user. 4.2
Here, we describe how we have used GRe ne to capture data transformation provenance events, how these events have been represented using our provenance models (described in section ??), and how we have made use of the IDT algebra (in section ??). We converted the contents of FilmAwardsForBestActress8 into a GRe ne project called \actresses".
We have applied our de nition of a Run (as stated in section ??) to actual implementations in the Usage Scenarios described below. Also here, we see the Google Re ne system as an implementation of an IDT, G (from De nition 1 of section ??).
Edit Usage Scenario If we want to change the entry in row 2 (of gure ??) from * "'1955 [[Meena Kumari]] "[[Parineeta (1953 lm)|Parineeta]]""' as "'Lolita" to \1935 John Wayne" and would like our system to keep a provenance record of this transaction, we can achieve that, in GRe ne, by viewing the column as a \Text Facet" and applying the GRe ne's \Edit" operation on row 2. Figure ?? shows us the before and after pictures of our Edit operation. The Events-to-RDF Statements Mapper automatically generates the RDF statements using the following mechanisms. A transformation function, e.g. \Edit", from GRe ne is automatically mapped to a type (an \rdf:type") of opmv:Process. What the operation gets mapped to in the Cogs ontology depends on the attribute of the data item that was the domain of the operation. For example, if the data item attribute is a Column, this gets mapped to a Cogs \ColumnOperation" class, while if the item attribute is a Row, this gets mapped to a Cogs \RowOperation" class. Since this is a transformation process, we made use of \cogs:TransformationProcess" class. voidp has a single class of ProvenanceEvent and every transformation operation is mapped to \voidp:ProvenanceEvent".
The system's Object-to-JavaClass Introspector is used to elicit the actual Java class responsible for the transformation. We have made use of a Cogs property, \cogs:programUsed", to specify the full path to this Java class. It allows the generated provenance data to communicate with the program semantics, in this way the intensional program semantics is linked up with the provenance extensional semantics. The data item that is actually made use of in the transformation process is mapped to \opmv:Artifact". To give us the process ow that is part of the transformation, we used two opmv properties: (a) \opmv:wasDerivedFrom", which tells us from which artifact this present data item is derived from, and (b) \opmv:wasGeneratedBy", which tells us from which process this data item is generated from. In order to store temporal information of when the derivation took place, we used \opmv:wasGeneratedAt", an opmv property that tells us the time of transformation.
The result of the transformation operation as RDF statements is below: @prefix id: <http://127.0.0.1:3333/project/1402144365904/> . id:MassCellChange-1092380975 rdf:type opmv:Process, cogs:ColumnOperation, cogs:TransformationProcess, voidp:ProvenanceEvent ; opmv:used <http://127.0.0.1:3333/project/1402144365904/
MassCellChange-1092380975/1_0> ; cogs:operationName "MassCellChange"^^xsd:string; cogs:programUsed "com.google.refine.operations.cell.
MassEditOperation"^^xsd:string; rdfs:label "Mass edit 1 cells in column ==List of winners=="^^xsd:string. <http://127.0.0.1:3333/project/1402144365904/MassCellChange-1092380975/1_0> rdf:type opmv:Artifact ; rdfs:label "* '''1955 [[Meena Kumari]]
'[[Parineeta (1953 film)|Parineeta]]''''' as '''Lolita'''"^^xsd:string. http://127.0.0.1:3333/project/1402144365904/MassCellChange-1092380975/1_1> rdf:type opmv:Artifact ; rdfs:label "* '''John Wayne'''"^^xsd:string; opmv:wasDerivedFrom <http://127.0.0.1:3333/project/1402144365904/
MassCellChange-1092380975/1_0>; opmv:wasGeneratedBy <http://127.0.0.1:3333/project/1402144365904/
MassCellChange-109238097
In the next section, we will describe the requirements we used for analysis. 5
Our approach and system were evaluated in relation to the set of requirements listed in [?][?] and enumerated below: 1. Decentralization: deployable one database at a time, without requiring co-operation among all databases at once. Coverage: High. Justi cation: Use of Semantic Web Standards and vocabularies (OPMV + Cogs + voidp) to reach a decentralized/interoperable provenance solution. 2. Data model independency: should work for data stored in at le, relational, XML, le system, Web site, etc., model. Coverage: High. Justi cation: Minimization of the interaction between the data level and the provenance level. Data is connected with its provenance descriptor by a provenance URI and the provenance representation is normalized as RDF/S. 3. Minimum impact to existing IDT practice: Provenance tracking is invisible to the user. Coverage: High. Justi cation: The provenance capture is transparent to the user. Provenance is captured by a lightweight instrumentation of the IDT platform, mapping program structures to the provenance model. 4. Scalability: to situations in which many databases cooperate to maintain provenance chain. Coverage: High. Justi cation: Usage of Semantic Web standards and vocabularies for provenance representation allows for the cooperation of multiple platforms for provenance management. 6
We categorize related work into two categories: (i) provenance management for manually curated data [?], and (ii) interactive data transformation models and tools [?][?]. Buneman et al. [?] propose a model for recording provenance of data in a single manually curated database. In their approach, the data is copied from external data sources or modi ed within the target database creating a copypaste model for describing user actions in assimilating external datasources into curated database records.
Raman and Hellerstein [?] describe Potters Wheel, an interactive data cleaning system which allowed users to specify transformations through graphic elements. Similarly, Wrangler [?] is an interactive tool based on the visual speci cation of data transformation. Both systems are similar to Google Re ne: [?], however, provides a more principled analysis of the interactive data transformation process, while [?] focuses on new techniques for specifying data transformations. The IDT and provenance model proposed in this work can be directly applied to both systems. 7
The world contains an unimaginably vast amount of data which is getting ever vaster ever more rapidly. These opportunities demand data transformation efforts for data analysis and re-purposing. Interactive data transformation (IDT) tools are becoming easily available to lower barriers to data transformation challenges. Some of these challenges will be solved by developing mechanisms useful for capturing the process ow and data lineage of these transformation processes in an interoperable manner. In this paper, we provide a formal model of IDT, a description of the design and architecture of an ontology-based provenance system used for mapping data transformation operations, and its implementation and validation on a popular IDT platform, Google Re ne. We shall be making our provenance extension to Google Re ne publicly accessible very soon. 8
This work was supported by the EnAKTing project, funded by EPSRC project number EP/G008493/1 and by the Science Foundation Ireland under Grant No. SFI/08/CE/I1380 (Lion-2).