A planner for exploratory data analysis
Robert St. Amant, Paul R. Cohen · 1997
Exploration often plays a central role in the early stages of scientific inquiry. One can rarely produce models of complex, unfamiliar phenomena on first contact with data. One must interpret suggestive features of the data, observe patterns these features indicate, and generate hypotheses to explain the patterns. Successive steps through the process can lead gradually to a better understanding of underlying structure in the data [Hoaglin et al., 1983; Good, 1983]. Exploratory data analysis (EDA) encompasses wide range of statistical tools [Tukey, 1977]. Simple exploratory results include histograms that describe discrete and continuous variables, schematic plots that give general characterizations of relationships, partitions of relationships that distinguish different modes of behavior, functional simplification of low-dimensionality relationships, and two-way tables such as contingency tables. These partial descriptions give different views of the data for a more complete, refined picture of underlying patterns. EDA techniques have found application across a variety of scientific domains. In well-known studies, researchers have used EDA to attack problems in grouping corporations [Chen et al., 1974], reducing TELSTAR data [Mallows, 1983], testing validity of approaches to ozone reduction [Cleveland et al., 1974], and examining disease characteristics [Diaconis, 1985]. Our own use of EDA has led us to a better understanding of complex AI systems [Cohen, 1995; St. Amant and Cohen, 1994]. Viewed as search, EDA poses a difficult problem. Suppose we define search operators to be the menu operations in a statistics package. We now have a range of flexible, powerful possibilities available: arithmetic composition of variables, such as those used in function finding; model-based variable decomposition, as performed by linear regression; partitioning and clustering operations, such as those used in numerical and conceptual clustering systems; feature extraction operations such as statistical summaries; various transfor-