Sparse learning under regularization framework

Regularization is a dominant theme in machine learning and statistics due to its prominent ability in providing an intuitive and principled tool for learning from high-dimensional data. As large-scale learning applications become popular, developing efficient algorithms and parsimonious models become promising and necessary for these applications. Aiming at solving large-scale learning problems, this thesis tackles the key research problems ranging from feature selection to learning with unlabeled data and learning data similarity representation. More specifically, we focus on the problems in three areas: online learning, semi-supervised learning, and multiple kernel learning.The first part of this thesis develops a novel online learning framework to solve group lasso and multi-task feature selection. To the best our knowledge, the proposed online learning framework is the first framework for the corresponding models. The main advantages of the online learning algorithms are that (1) they can work on the applications where training data appear sequentially; consequently, the training procedure can be started at any time; (2) they can handle data up to any size with any number of features. The efficiency of the algorithms is attained because we derive closed-form solutions to update the weights of the corresponding models. At each iteration, the online learning algorithms just need O (d) time complexity and memory cost for group lasso, while they need O (d × Q) for multi-task feature selection, where d is the number of dimensions and Q is the number of tasks. Moreover, we provide theoretical analysis for the average regret of the online learning algorithms, which also guarantees the convergence rate of the algorithms. In addition, we extend the online learning framework to solve several related models which yield more sparse solutions.The second part of this thesis addresses a general scenario of semi-supervised learning for the binary classification problern, where the unlabeled data may be a mixture of relevant and irrelevant data to the target binary classification task. Without specifying the relatedness in the unlabeled data, we develop a novel maximum margin classifier, named the tri-class support vector machine (3C-SVM), to seek an inductive rule that can separate these data into three categories: –1, +1, or 0. This is achieved by adopting a novel min loss function and following the maximum entropy principle. For the implementation, we approximate the problem and solve it by a standard concaveconvex procedure (CCCP). The approach is very efficient and it is possible to solve large-scale datasets.The third part of this thesis focuses on multiple kernel learning (MKL) to solve the insufficiency of the L1-MKL and the Lp-MKL models. Hence, we propose a generalized MKL (GMKL) model by introducing an elastic net-type constraint on the kernel weights. More specifically, it is an MKL model with a constraint on a linear combination of the L1-norm and the square of the L2-norm on the kernel weights to seek the optimal kernel combination weights. Therefore, previous MKL problems based on the L1-norm or the L2-norm constraints can be regarded as its special cases. Moreover, our GMKL enjoys the favorable sparsity property on the solution and also facilitates the grouping effect. In addition, the optimization of our GMKL is a convex optimization problem, where a local solution is the globally optimal solution. We further derive the level method to efficiently solve the optimization problem.