Improving Random Forests by Correlation-Enhancing Projections and Sample-Based Sparse Discriminant Selection

Random Forests (RF) is a learning technique with very low run-time complexity. It has found a niche application in situations where input data is low-dimensional and computational performance is paramount. We wish to make RFs more useful for high dimensional problems, and to this end, we propose two extensions to RFs: Firstly, a feature selection mechanism called correlation-enhancing projections, and secondly sparse discriminant selection schemes for better accuracy and faster training. We evaluate the proposed extensions by performing age and gender estimation on the MORPH-II dataset, and demonstrate near-equal or improved estimation performance when using these extensions despite a seventy-fold reduction in the number of data dimensions.