Testing the ability of the classifier development method to work well for a dataset with very many, but few useful, features: predicting ten year survival for patients with prostate cancer

This investigation used the same datasets as above with the same goal to predict 10-year survival. Here we compared the DRC classifier approach with the RF. To mimic the situation of very many features, with only a few with utility for the problem in question, we added 10,000 randomly generated gaussian features (mean = 0 and standard deviation = 1) to both the development and validation data sets. For the DRC approach, rank-based kNNs were used as atomic classifiers, to avoid any problems with differences in scale between the original and randomly generated features. All kNN classifiers (k = 7) using the 10,343 features singly and pairs of features that passed single feature filtering were considered. Filtering was set as in the previous problem and resulted in around 25% of atomic classifiers considered passing filtering and 100,000 dropout iterations were used.

DRC and RF were generated using identical training/test set realizations for 9 subsets each of the development set with N = 24, 48, 60, 72, 84, 93, and 105 samples per class. All other parameters used were the same as listed above.