A0427
Title: Feature subset sampling for prediction in high-dimensional linear models
Authors: Hua Liang - George Washington University (United States) [presenting]
Abstract: The high-dimensional prediction problem without the sparsity assumption is considered and a feature subset sampling (FSS) method is introduced, which chooses features by random sampling to get a feature subset, then does the prediction based on the sampled feature subset. This strategy greatly reduces the dimension and saves computational time. We explore this strategy under the ridge regression framework and derive the risk bound as a statistical guarantee of the method. We suggest two sampling strategies: uniform sampling and sampling proportional to the feature information score. Finally, we present detailed empirical studies for illustrating the numerical performance of the method. We demonstrate that the prediction risk of the FSS method can close to that of the full features under the mild assumptions.
