ITR-AP: A Program for Predicting and Understanding Data

University Of California-Berkeley, Berkeley CA

Investigators

Abstract

ABSTRACT DMS 0112734- The two main uses of much of the scientific data currently being collected are first to use the data to predict the types of future objects observed by means of computerized algorithms. For instance, one such problem is to develop algorithms that will classify the millions of stellar objects recorded on photographic images by optical and radio telescopes. Second: to understand which variables are discriminating between different types of objects. An example is in locating the gene activity that discriminates between cancerous and non cancerous DNA. This project builds on a recently discovered algorithm called random forests which, when further developed and combined with interactive graphical displays, will provide an advanced tool for answering these questions. Random forests is a new prediction algorithm coming from the Machine Learning context that functions by combining hundreds of randomly generated binary decision. It has demonstrated state-of-the-art prediction accuracy on large data sets with thousands of variables. It generates a wealth of information about the data other than the prediction. This information can be used to estimate variable importance, clustering, density estimation etc. To make this information more readily understandable to the user, interactive graphics such as parallel coordinates and hierarchical cluster diagrams linked together will be incorporated into the design of a program using a more highly developed random forests algorithm as the underlying engine.