Learning Non-Linear Feature Maps
This work addresses a computational bottleneck in feature selection for domains like biology, offering a more scalable solution.
The paper tackles the computational inefficiency of non-linear feature selection methods for high-dimensional datasets by proposing randSel, a randomized algorithm with strong probabilistic guarantees for identifying relevant features, achieving better performance than competitive approaches in experiments.
Feature selection plays a pivotal role in learning, particularly in areas were parsimonious features can provide insight into the underlying process, such as biology. Recent approaches for non-linear feature selection employing greedy optimisation of Centred Kernel Target Alignment(KTA), while exhibiting strong results in terms of generalisation accuracy and sparsity, can become computationally prohibitive for high-dimensional datasets. We propose randSel, a randomised feature selection algorithm, with attractive scaling properties. Our theoretical analysis of randSel provides strong probabilistic guarantees for the correct identification of relevant features. Experimental results on real and artificial data, show that the method successfully identifies effective features, performing better than a number of competitive approaches.