Technical Sessions

We describe ReproZip, a tool that makes it easier for authors to publish reproducible results and for reviewers to validate these results. By tracking operating system calls, ReproZip systematically captures detailed provenance of existing experiments, including data dependencies, libraries used, and configuration parameters. This information is combined into a package that can be installed and run on a different environment. An important goal that we have for ReproZip is usability. Besides simplifying the creation of reproducible results, the system also helps reviewers. Because the package is self contained, reviewers need not install any additional software to run the experiments. In addition, ReproZip generates a workflow specification for the experiment. This not only enables reviewers to execute this specification within a workflow system to explore the experiment and try different configurations, but also the provenance kept by the workflow system can facilitate communication between reviewers and authors.

This paper presents a provenance-based technique to support undoing and redoing data analysis tasks. Our technique targets scientists who experiment with combinations of approaches to processing raw data into presentable datasets. Raw data may be noisy and in need of cleaning, it may suffer from sensor drift that requires retrospective calibration and data correction, or it mayneed gap-filling due to sensor malfunction or environmental conditions. Different raw datasets may have different issues requiring different kinds of adjustments, and each issue may potentially be handled by different approaches. Thus, scientists must often experiment with different sequences of approaches. In our work, we show how provenance information can be used to facilitate this kind of experimentation with scientific datasets. We describe an approach that supports the ability to (1) undo a set of tasks while setting aside the artifacts and consequences of performing those tasks, (2) replace, remove, or add a data-processing technique, and (3) redo automatically those set aside tasks that are consistent with changed technique. We have implemented our technique and demonstrate its utility with a case study of a common, sensor-network, data-processing scenario showing how our approach can reduce the cost of changing intermediate data-processing techniques in a complex, data-intensive process.

Data mining aims at extracting useful information from large datasets. Most data mining approaches reduce the input data to produce a smaller output summarizing the mining result. While the purpose of data mining (extracting information) necessitates this reduction in size, the loss of information it entails can be problematic. Specifically, the results of data mining may be more confusing than insightful, if the user is not able to understand on which input data they are based and how they were created. In this paper, we argue that the user needs access to the provenance of mining results. Provenance, while extensively studied by the database, workflow, and distributed systems communities, has not yet been considered for data mining. We analyze the differences between database, workflow, and data mining provenance, suggest new types of provenance, and identify new use cases for provenance in data mining. To illustrate our ideas, we present a more detailed discussion of these concepts for two typical data mining algorithms: frequent item set mining and multi-dimensional scaling.

Systems that gather fine-grained provenance metadata must process and store large amounts of information. Filtering this metadata as it is collected has a number of benefits, including reducing the amount of persistent storage required and simplifying subsequent provenance queries. However, writing these filters in a procedural language is verbose and error prone. We propose a simple declarative language for processing provenance metadata and evaluate it by translating filters implemented in SPADE, an open-source provenance collection platform.

This paper presents an extension to the W3C PROV provenance model, aimed at representing process structure. Although the modelling of process structure is out of the scope of the PROV specification, it is beneficial when capturing and analyzing the provenance of data that is produced by programs or other formally encoded processes. In the paper, we motivate the need for such and extended model in the context of an ongoing large data federation and preservation project, DataONE, where provenance traces of scientific workflow runs are captured and stored alongside the data products. We introduce new provenance relations for modelling process structure along with their usage patterns, and present sample queries that demonstrate their benefit.

We introduce HadoopProv, a modified version of Hadoopthat implements provenance capture and analysis in MapReduce jobs. It is designed to minimise provenancecapture overheads by (i) treating provenance tracking in Map and Reduce phases separately, and (ii) deferringconstruction of the provenance graph to the query stage. Provenance graphs are later joined on matching intermediate keys of the Map and Reduce provenance files. In our prototype implementation, HadoopProv has an overhead below 10% on typical job runtime (<7% and <30% average temporal increase on Map and Reduce tasks respectively). Additionally, we demonstrate that provenance queries are serviceable in O (k log n), where n is the number of records per Map task and k is the set of Map tasks in which the key appears.