An Introspective Comparison of Random Forest-Based Classifiers for the Analysis of Cluster-Correlated Data by Way of RF++
Karpievitch, YV and Hill, EG and Leclerc, AP and Dabney, AR and Almeida, JS, An Introspective Comparison of Random Forest-Based Classifiers for the Analysis of Cluster-Correlated Data by Way of RF++, PLoS One, 4, (9) Article 7087. ISSN 1932-6203 (2009) [Refereed Article]
Many mass spectrometry-based studies, as well as other biological experiments produce cluster-correlated data. Failure to account for correlation among observations may result in a classification algorithm overfitting the training data and producing overoptimistic estimated error rates and may make subsequent classifications unreliable. Current common practice for dealing with replicated data is to average each subject replicate sample set, reducing the dataset size and incurring loss of information. In this manuscript we compare three approaches to dealing with cluster-correlated data: unmodified Breiman's Random Forest (URF), forest grown using subject-level averages (SLA), and RF++ with subject-level bootstrapping (SLB). RF++, a novel Random Forest-based algorithm implemented in C++, handles cluster-correlated data through a modification of the original resampling algorithm and accommodates subject-level classification. Subject-level bootstrapping is an alternative sampling method that obviates the need to average or otherwise reduce each set of replicates to a single independent sample. Our experiments show nearly identical median classification and variable selection accuracy for SLB forests and URF forests when applied to both simulated and real datasets. However, the run-time estimated error rate was severely underestimated for URF forests. Predictably, SLA forests were found to be more severely affected by the reduction in sample size which led to poorer classification and variable selection accuracy. Perhaps most importantly our results suggest that it is reasonable to utilize URF for the analysis of cluster-correlated data. Two caveats should be noted: first, correct classification error rates must be obtained using a separate test dataset, and second, an additional post-processing step is required to obtain subject-level classifications. RF++ is shown to be an effective alternative for classifying both clustered and non-clustered data. Source code and stand-alone compiled versions of command-line and easy-to-use graphical user interface (GUI) versions of RF++ for Windows and Linux as well as a user manual (Supplementary File S2) are available for download at: http://sourceforge.org/projects/rfpp/ under the GNU public license.