SVM Kernel Optimization: An Example in Yeast Protein Subcellular Localization Prediction

Localization of proteins, a flourishing area in bioinformatics, can help us understand their respective functions. Currently there exist a number of localization approaches based on machine learning algorithms, and support vector machines (SVMs) have been used extensively. However, in terms of kernel optimization, a critical step in SVM design, there is no well-established systematic method so far. In this paper, we apply the Levenberg-Marquardt (LM) algorithm in kernel optimization, an algorithm that we believe will have better optimization performance than simple gradient descent and that hasn’t been used in this field according to our knowledge, and test it using protein location data from yeast. We then experimentally compare the performance of our optimized system with others. Results show that automated parameter optimization can improve classification accuracy. Further research can complement such a method by controlling over-fitting.