Efficiently Learning Nonlinear Classifiers for Domain Specific Performance Measures

In practical applications, machine learning algorithms are often needed to learn classifiers that optimize domain specific performance measures. In the past, the research has focused on learning the needed classifier in isolation, yet learning nonlinear classifier for nonlinear and nonsmooth performance measures is still hard. In this paper, rather than learning the needed classifier by optimizing the concerned performance measure directly, we propose a new framework ELperf to circumvent this problem. In ELperf, we first train auxiliary classifiers by optimizing an easy-to-handle performance measure; and then adapt these auxiliary classifiers to optimize the concerned performance measure. Under the function-level adaptation framework, the classifier adaptation problem of ELperf is formulated as a quadratic programming problem, which is similar to linear SVMperf and can be efficiently solved. Practically, by using nonlinear auxiliary classifiers, ELperf can generate nonlinear classifier that optimizes required performance measure, whilst keeping computational efficiency. In extensive empirical studies, we show that training auxiliary classifiers using accuracy is sufficiently good for ELperf, and other performance measure is not necessary. As well, it is shown that ELperf is effective and efficient in training classifiers that optimize performance measures, and even its classifier adaptation procedure is more efficient than linear SVMperf.