Adaptive Dual Greedy: Using an LTF evaluation algorithm to reduce the cost of using SVM

During test-time, SVM evaluation can be expensive if the features are time-consuming to extract. This work uses an approximation algorithm for the submodular set cover problem to select a subset of all features that may be used during test-time to arrive at the same accuracy as if all features were used. On synthetic and UCI datasets this approach shows reductions in prediction cost, while maintaining prediction accuracy.