Mining scientific data archives through metadata generation

Data analysis and management tools typically have not supported the documenting of data, so scientists must manually maintain all information pertaining to the context and history of their work. This metadata is critical to effective retrieval and use of the masses of archived data, yet little of it exists on-line or in an accessible format. Exploration of archived legacy data typically proceeds as a laborious process, using commands to navigate through file structures on several machines. This file-at-a-time approach needs to be replaced with a model that represents data as collections of interrelated objects. The tools that support this model must focus attention on data while hiding the complexity of the computational environment. This problem was addressed by developing a tool for exploring large amounts of data in UNIX directories via automatic generation of metadata summaries. This paper describes the model for metadata summaries of collections and the Data Miner tool for interactively traversing directories and automatically generating metadata that serves as a quick overview and index to the archived data. The summaries include thumbnail images as well as links to the data, related directories, and other metadata. Users may personalize the metadata by adding a title and abstract to the summary, which is presented as an HTML page viewed with a World Wide Web browser. We have designed summaries for 3 types of collections of data: contents of a single directory; virtual directories that represent relations between scattered files; and groups of related calculation files. By focusing on the scientists` view of the data mining task, we have developed techniques that assist in the ``detective work `` of mining without requiring knowledge of mundane details about formats and commands. Experiences in working with scientists to design these tools are recounted