Ranked Similarity Search of Scientific Datasets: An Information Retrieval Approach

Abstract

In the past decade, the amount of scientific data collected and generated by scientists has grown dramatically. This growth has intensified an existing problem: in large archives consisting of datasets stored in many files, formats and locations, how can scientists find data relevant to their research interests? We approach this problem in a new way: by adapting Information Retrieval techniques, developed for searching text documents, into the world of (primarily numeric) scientific data. We propose an approach that uses a blend of automated and curated methods to extract metadata from large repositories of scientific data. We then perform searches over this metadata, returning results ranked by similarity to the search criteria. We present a model of this approach, and describe a specific implementation thereof performed at an ocean-observatory data archive and now running in production. Our prototype implements scanners that extract metadata from datasets that contain different kinds of environmental observations, and a search engine with a candidate similarity measure for comparing a set of search terms to the extracted metadata. We evaluate the utility of the prototype by performing two user studies; these studies show that the approach resonates with users, and that our proposed similarity measure performs well when analyzed using standard Information Retrieval evaluation methods. We performed performance tests to explore how continued archive growth will affect our goal of interactive response, developed and applied techniques that mitigate the effects of that growth, and show that the techniques are effective. Lastly, we describe some of the research needed to extend this initial work into a true "Google for data".

Demonstrating "Data Near Here"

“…The per-ter a field as a lightw n. The per-term ance function, w the search term data bounds dec n is not symmetr using the bound e data varies uni n with this assum lting ranked list when using the ent it would rece e underlying data ted individually ed over the datas dity of the assu parisons of the tiple scores at a same values. T d behind the oth our similarity fu data within the uch summarizati search engine ty between the g [14]. The result ity [14,15] We separately t poral, and envi es from a prior the dataset con that study.…”

“…T d behind the oth our similarity fu data within the uch summarizati search engine ty between the g [14]. The result ity [14,15] We separately t poral, and envi es from a prior the dataset con that study. We s Figure 5: in eac oint-wise calcul culated from the scores for tempo atial search term in Figure 5c.…”

“…Geospatially, many of the datasets are either from a stationary device (in which case the result will be the same from both methods), or from a track from a mobile device. For mobile devices, the similarity measure is based on the closest and farthest points of a simplified line representing the locations of the observations [14], so it is less obvious that the scores would be close for these datasets. The environmental variables themselves generally shift relatively slowly, within some kind of limits or in a pattern.…”