Learning and optimization of 'fhe kernel functions from insufficiently labeled data

Loading...
Thumbnail Image
Date
2010
Authors
Abbasnejad, M.Ehsan
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Amongst all the machine learning techniques, kernel methods are increasingly becoming popular due to their efficiency, accuracy and ability to handle high-dimensional data. The fundamental problem related to these learning techniques is the selection of the kernel function. Therefore, learning the kernel as a procedure in which the kernel function is selected for a particular dataset is highly important. In this thesis, two approaches to learn the kernel function are proposed: transferred learning of the kernel and an unsupervised approach to learn the kernel. The first approach uses transferred knowledge from unlabeled data to cope with situations where training examples are scarce. Unlabeled data is used in conjunction with labeled data to construct an optimized kernel using Fisher discriminant analysis and maximum mean discrepancy. The accuracy of classification which indicates the number of correctly predicted test examples from the base kernels and the optimized kernel are compared in two datasets involving satellite images and synthetic data where proposed approach produces better results. The second approach is an unsupervised method to learn a linear combination of kernel functions. Here, the global intrinsic structure of the unlabeled data is inferred through a measure called influence, which is computed by constructing a weighted graph and performing a random walk upon it. The measure of influence in the feature space is probabilistically related to the input space that yields an optimization problem to be solved. The optimization problem is formulated in two different convex settings, namely linear and semidefinite programming, depending on the type of kernel combination considered. Here, the contributions are twofold: first, a novel unsupervised approach to learn the kernel function, and second, a method to infer the local similarity represented by the kernel function by measuring the global influence of each point towards the structure of the dataset. The proposed approach focuses on the kernel selection which is independent of the kernel-based learning algorithm. The empirical evaluation of the proposed approach on image and text classification shows the effectiveness of the algorithm in obtaining more accurate results.
Description
Keywords
Kernel functions , Labeled data
Citation