ITR/SMALL/Scientific Frontiers: Task-Specific Data Reduction and Mining in Spatial-Temporal Domains

Temple University, Philadelphia PA

Investigators

Abstract

Advances in earth sciences have spurred a large growth in spatial- temporal databases with multidimensional attributes. Unfortunately, the improvement in data collection techniques has not been accompanied with the efficient methods that analyze large databases, which currently limits their use. To address the need for facilitating large-scale knowledge discovery at a wide range of spatial-temporal datasets the objective of the proposed project is development of novel techniques for (1) task-specific data reduction in spatial-temporal databases; and (2) efficient knowledge discovery on reduced spatial-temporal data. The first task builds on the observation that the amount of collected data is not necessarily correlated with specific knowledge discovery tasks, and that real-life datasets are often highly redundant. The second task addresses a number of open questions for more efficient mining of reduced spatial-temporal datasets including exploring trade-offs between model complexity and data reduction for efficient learning, exploiting residual correlations for prediction, and task-specific partitioning of nonstationary spatial-temporal data. The proposed techniques will be evaluated on multi-source datasets with various complexities. This project will result in guidelines for effective dissemination of spatial-temporal data sets to wider research audience and algorithmic tools for more successful knowledge discovery from such data.