Support Vector Machine to synthesise kernels
Suppose we are given two sets of features from distinct sources about objects that need to be classified. The question we wish to answer is how to combine them into one classification rule, which can outperform a classifier based on one feature set only. To aggregate distinct single features into one efficient classifier boosting type methods based on a relatively simple, convex combination of each feature have proved fruitful. We introduce a new method, SVM_2k, which amalgamates the capabilities of the SVM and KCCA to give a more sophisticated combination rule than boosting allows. The SVM provides the tools for classification given distinct feature sets and a KCCA like method achieves the synthesis of the learners. We use the distance minimising version of KCCA to unite the different SVM problems. We construct the learner via a unified optimisation problem to create a consistent learning rule and solve this problem using a version of the Augmented Lagrangian Method.