Using Random Forest Proximities for Analyzing Tissue Microarray Data

• Activity Number: 61
• Type: Contributed
• Date/Time: Sunday, August 3, 2003 : 4:00 PM to 5:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Tao Shi*+ and Zixing Fang and Steve Horvath
• Companies: University of California, Los Angeles and University of California, Los Angeles and University of California, Los Angeles
• Keywords: tissue microarray ; random forest ; proximity measures ; multidimensional scaling ; dissimilarity measure

Abstract

A random forest (L. Breiman 1999) predictor is an ensemble of individual classification tree predictors. Interestingly, the RF prediction method can also be used to arrive at an "internal" proximity measure between observations even when no outcome has been specified. This proximity measure can be used for clustering or for creating multidimensional scaling plots. To define the proximity measure one needs to generate a set of synthetic data from a null distribution corresponding to no clusters in the data set. Apart from choosing a way of generating synthetic observations, one needs to specify the number of random features mtry. We will discuss several properties of RF clustering. First, we describe the nature of cluster produced by RF clustering. Second, we recommend guidelines for choosing the mtry parameter. Third, we will compare different multidimensional scaling methods. Fourth, we will contrast standard distance measures, e.g., Euclidean, to the RF dissimilarity measure. We apply RF clustering to the domain of tissue microarray data where we find that the RF dissimilarity measure is superior to a Euclidean distance.