Semi-Supervised Learning With Multiple Views

Learning from examples is an important ability of human beings. The goals of machine learning is to simulate the learning process of human. By applying the research results of neurophysiology and cognitive psychology to construct the computational models and algorithms, machine learning aims to predict the unseen examples, which is an important part of artificial intelligence and neural computing.With the development of information technology, there are abundant unlabeled examples while the number of labeled examples is limited, because labeling the examples requires human efforts. The word "label" indicates the desired output of the example, e.g. in classification it indicates the category of the example. Traditional supervised learning needs a large number of labeled examples to construct the model, which has poor performance when the label is scare. So, semi-supervised learning which exploits unlabeled examples in addition to labeled examples to improve learning performance has been a hot topic recently.Many problems in machine learning involve examples that are naturally comprised of multiple views. In this dissertation, several key problems of exploiting multiple views to effectively learn from labeled and unlabeled examples are studied, which include the theory and the algorithm of multi-view semi-supervised learning, the construction of multiple views, and the combination of multi-view semi-supervised learning with active learning. The methods and technologies proposed in this dissertation are verified through sufficient experiments.The main contributions of this dissertation are summarized as follows:1. We propose a new regularization method in multi-view semi-supervised learning. Learning from limited examples is an ill-posed inverse problem, to which regularization method has to be used. By exploiting the metric structure of the hypotheses space, we define the smoothness and consistency of a hypothesis. A two levels regularization algorithm is presented which uses the smoothness to regularize the within-view learning process while uses the consistency to regularize the between-view learning process. The prediction error of the algorithm is analyzed. Encouraging experimental results are presented on both synthetic and real world datasets.2. We propose a new graph-based multi-view semi-supervise learning method. As graph can be used to represent the examples and the relationship between examples, multiple graphs can be used to represent multi-view examples. By extending graph-based semi-supervise learning to solve the multi-view learning problem, a semi-supervised learning algorithm with multi-graph is presented, which using unlabeled examples to learning in each graph while using unlabeled examples to co-learning between graphs. The experimental results on real world dataset show that our method is more accurate comparing with graph-based single-view semi-supervised learning methods.3. We propose a new multi-view construction method. By projecting examples into the random subspaces of the feature space, we construct views of the original examples. A multi-view semi-supervised learning algorithm is presented, which trains a classifier in each view and chooses the most confident examples of each classifier to train the other classifiers. Random discrimination theory is used to analyze the performance of the algorithm. The experimental results on real world datasets show that our method is effective when the feature are abundant.4. We propose a new active learning method and combine it with the multi-view semi-supervised learning method. When the learner can interact with the environment, it can choose some examples to query their labels from the user. By selecting the example nearest to the classification hyperplane, we present an active learning algorithm which ask the user to label the least confident examples of the learner. Then, we incorporate the active learning process into the multi-view semi-supervised learning process. For each view, the most confident examples are selected to enlarge the training set of the other view, while the least confident examples are selected to query. The experimental results on both synthetic and real world datasets demonstrate that the classification performance can be improved distinctly with the proposed active learning method.