More than a decade ago, Peter Buneman used the term curated databases to refer to databases that are created and maintained using the (often substantial) effort and domain expertise of humans. These human experts clean the data, integrate it with new sources, prepare it for analysis, and share the data with other experts in their field. In data curation, one seeks to support human curators in all activities needed for maintaining and enhancing the value of their data over time. Curation includes data provenance, the process of understanding the origins of data, how it was created, cleaned, or integrated. Big Data offers opportunities to solve curation problems in new ways. The availability of massive data is making it possible to infer semantic connections among data, connections that are central to solving difficult integration, cleaning, and analysis problems. Some of the nuanced semantic differences that eluded enterprise-scale curation solutions can now be understood using evidence from Big Data. Big Data Curation leverages the human expertise that has been embedded in Big Data, be it in general knowledge data that has been created through mass collaboration, or in specialized knowledge-bases created by incentivized user communities who value the creation and maintenance of high quality data. In this talk, I describe our experience in Big Data Curation. This includes our experience over the last five years curating NIH Clinical Trials data that we have published as Open Linked Data at linkedCT.org. I overview how we have adapted some of the traditional solutions for data curation to account for (and take advantage of) Big Data.

The YesWork ow (YW) toolkit aims to provide users of scripting languages such as Python, Perl, and R with many of the benefits of scientific workflow automation. YW requires neither the use of a workflow engine nor the overhead of adapting or instrumenting code to run in such a system. Instead, YW enables scientists to annotate their scripts with special comments that reveal the main computational blocks and dataflow dependencies otherwise implicit in scripts. YW tools extract and analyze these comments, represent scripts in terms of entities based on a typical scientific workflow model, and provide graphical workflow views (i.e., prospective provenance) of scripts. In this paper, we present a new extension of YW for inferring retrospective provenance from script executions without relying on a runtime provenance recorder. Instead we exploit the common practice of scientists to embed important pieces of provenance in directory structures and file names. For such “provenance-friendly” data organizations, we offer a new annotation mechanism based on URI templates. YW uses these to link conceptual-level prospective provenance with data files created at runtime, resulting in a powerful, integrated model of prospective and retrospective provenance.We present scientifically meaningful retrospective provenance queries for investigating an execution of a data acquisition workflow implemented as a Python script, and show how these queries can be evaluated using the YW toolkit.

Capturing provenance usually involves the direct observation and instrumentation of the execution of a program or workflow. However, this approach restricts provenance analysis to pre-determined programs and methods. This may not pose a problem when one is interested in the provenance of a well-defined workflow, but may limit the analysis of unstructured processes such as interactive desktop computing. In this paper, we present a new approach to capturing provenance based on full execution record and replay. Our approach leverages full-system execution trace logging and replay, which allows the complete decoupling of analysis from the original execution. This enables the selective analysis of the execution using progressively heavier instrumentation.

The so-called big data revolution has been characterised by an increase in sources of data as well as in the volume of data to be processed.

In many cases—for example, network behaviour and control, security monitoring, enterprise management information—the data for situation awareness and decision-making is drawn from multiple sources and must be integrated into a coherent whole as far as possible.

This process generally requires both machines and human analysts and experts, It includes compensating for different formats, different granularities and resolutions, identifying and correcting errors (both systematic and intermittent), as well as managing uncertainties and gaps in data. Often the process requires assumptions and choices to be made in arriving at a reasonably robust overview of a situation—for example, in deciding that a failed attempt to access a building is potentially malicious, we might need to take account of someone's recent travel, long term patterns of behaviour, current schedules of close colleagues, etc. where each of these components may have been derived from lower-level raw data. Provenance in this context refers to the derivation pathways and their overall reliability.

In this talk, I will describe the use of graded (fuzzy) representations in modelling and managing the uncertainties, reliability and granularity of derived data in combining sources for situation awareness.

Organizations and teams collect and acquire data from various sources, such as social interactions, financial transactions, sensor data, and genome sequencers. Different teams in an organization as well as different data scientists within a team are interested in extracting a variety of insights which requires combining and collaboratively analyzing datasets in diverse ways. DataHub is a system that aims to provide robust version control and provenance management for such a scenario. To be truly useful for collaborative data science, one also needs the ability to specify queries and analysis tasks over the versioning and the provenance information in a unified manner. In this paper, we present an initial design of our query language, called VQuel, that aims to support such unified querying over both types of information, as well as the intermediate and final results of analyses. We also discuss some of the key language design and implementation challenges moving forward.

When performing automatic provenance collection within the operating system, inevitable storage overheads are made worse by the fact that much of the generated lineage is uninteresting, describing noise and background activities that lie outside the scope the system’s intended use. In this work, we propose a novel approach to policy-based provenance pruning—leverage the confinement properties provided by Mandatory Access Control (MAC) systems in order to identify subdomains of system activity for which to collect provenance. We consider the assurances of completeness that such a system could provide by sketching algorithms that reconcile provenance graphs with the information flows permitted by the MAC policy. We go on to identify the design challenges in implementing such a mechanism. In a simplified experiment, we demonstrate that adding a policy component to the Hi-Fi provenance monitor could reduce storage overhead by as much as 82%. To our knowledge, this is the first practical policy-based provenance monitor to be proposed in the literature.

Provenance databases are an important asset in data analytics of large-scale scientific data. The data derivation path allows for identifying parameters, files and domain data values of interest. In scientific workflows, provenance data is automatically captured by workflow systems. However, the power of provenance data analyses depends on the expressiveness of domain-specific data along the provenance traces. While much has been done through the W3C PROV initiative and its PROV-DM to represent generic provenance data, representing domain-specific data in provenance traces has received little attention, yet it accounts for a large number of provenance analytical queries. Such queries are based on selections on data values from input/output artifacts along workflow activities. There are several problems in modeling and capturing values from domain-specific attributes, some of them are related to managing provenance granularity, others to addressing data values hidden inside files and representing the semantics of domain data. In this work, we discuss these open issues and propose some alternatives to domain-specific provenance data capture, representation, storage and queries. Addressing these issues may be decisive in using provenance to drive scientific data analyses at large-scale.

Scripting languages like Python, R, andMATLAB have seen significant use across a variety of scientific domains. To assist scientists in the analysis of script executions, a number of mechanisms, e.g., noWorkflow, have been recently proposed to capture the provenance of script executions. The provenance information recorded can be used, e.g., to trace the lineage of a particular result by identifying the data inputs and the processing steps that were used to produce it. By and large, the provenance information captured for scripts is fine-grained in the sense that it captures data dependencies at the level of script statement, and do so for every variable within the script. While useful, the amount of recorded provenance information can be overwhelming for users and cumbersome to use. This suggests the need for abstraction mechanisms that focus attention on specific parts of provenance relevant for analyses. Toward this goal, we propose that fine-grained provenance information recorded as the result of script execution can be abstracted using user-specified, workflow-like views. Specifically, we show how the provenance traces recorded by noWorkflow can be mapped to the workflow specifications generated by YesWorkflow from scripts based on user annotations. We examine the issues in constructing a successful mapping, provide an initial implementation of our solution, and present competency queries illustrating how a workflow view generated from the script can be used to explore the provenance recorded during script execution.

Today’s programming languages provide no support for data provenance. In a world that increasingly relies on data, we need provenance to judge the reliability of data and therefore should aim for making it easily accessible to programmers. We report our work in progress on an extension to the Links programming language that builds on its support for language-integrated query to support where-provenance queries through query rewriting and a type system extension that distinguishes provenance metadata from other data. Our approach aims to work solely within the language implementation and thus require no changes to the database system. The type system together with automatic propagation of provenance metadata will prevent programmers from accidentally changing provenance, losing it, or misattributing it to other data.

In recent work, Salimi and Bertossi provide a tight connection between causality and tuple-based data repairs. We investigate this connection between causality and two other kinds of repair models. First, we consider cell-based V-repairs, i.e., repairs that are obtained by modifying cells in the data. In contrast, tuple-based repairs only allow for the deletion of tuples. Second, we introduce a new notion of repairs, called chase repairs, that take into account the procedural (chase) steps that lead to a repair. We establish a connection between causes (and the associated notion of responsibility) and V-repairs, and analyse the complexity of verifying whether a cell is a cause and whether its responsibility is above a certain threshold. Our understanding of chase repairs is still very preliminary, and we argue that provenance models that are specifically targeted to data repairs and data quality in general are needed to make formal connections between causality and chase repairs.