Automated Data Discovery in Similarity Score Queries

A vast amount of information is being stored in scientific databases on the web. The dynamic nature of the scientific data, the cost of providing an up-to-date snapshot of the whole database, and proprietary considerations compel the database owners to hide the original data behind search interfaces. The information is often provided to researchers through similarity-search query interfaces, which limits a proper and focused analysis of the data. In this study, we present systematic methods of data discovery through similarity-score queries in such “uncooperative” databases. The methods are generalized to multidimensional data, and to L-p norm distance functions. The accuracy and performance of our methods are demonstrated on synthetic and real-life datasets. The methods developed in this study enable the scientists to obtain the data within the range of their research interests, overcoming the limitations of the similarity-search interface. The results of this study also present implications in data privacy and security areas, where the discovery of the original data is not desired.