Partitions and Piecewise Models
Tamraparni Dasu, Theodore J. Johnson · Wiley series in probability and statistics · 2003
In this chapter, we discuss methods for dividing up the data into manageable pieces for the purpose of data summarization, data cleaning, and scaling up analyses to massive data through piecewise models built using EDM summaries. We discuss linear partitions and nonlinear partitions. DataSpheres, a particular type of nonlinear partitions are also discussed. We describe set comparison techniques based on EDM summaries of partitions for detecting changes in data sets, whether caused by genuine shifts in distribution or by data glitches. Next, we describe the process of approximating complex structure in data using a collection of simpler models. We give instances of such approximations, namely piecewise linear regression and one-pass classification.