There is a rapidly widening gap between our ability to collect data and our ability to explore, analyze, and understand the data. As a result, useful information is overlooked, and the potential benefits of increased computational and data gathering capabilities only partially realized. This problem of data overload is becoming a serious impediment to scientific advancement in areas as diverse as counter-proliferation, the Accelerated Strategic Computing Initiative (ASCI), astrophysics, computer security, and climate modeling, where vast amounts of data are collected through observations or simulations. To improve the way in which scientists extract useful information from their data, we are developing a new generation of tools and techniques based on data mining. Data mining is the semi-automated discovery of patterns, associations, anomalies, and statistically significant structures in data. It consists of two steps--in data pre-processing, we extract high-level features from the data, and in pattern recognition, we use the features to identify and characterize patterns in the data. In this project, our focus is on developing scalable algorithms for the pattern recognition task of classification. Our goal is to improve the performance of these algorithms, without sacrificing accuracy. We are demonstrating these techniques using an astronomy application, namely the detection of radio-emitting galaxies with a bent-double morphology in the FIRST survey. Our research has been incorporated into software to make it easily accessible to LLNL scientists. The author describes their accomplishments in each of these three areas.
