Learning in semi-supervised model and sparse model

Machine learning has the ability to learn from data automatically and improve the decision-making process with experience. Among different types of machine learning paradigms, semi-supervised learning (SSL) has recently attracted growing attention in the community and also been used to a wide range of real-world applications. In the semi-supervised learning paradigm, how to obtain informative sub-feature sets is important to speed up the learning process. We utilize the max-Relevance and min-Redundancy criteria and propose a new semi-supervised feature selection method. Meanwhile, we use the diversity to propose a new co-training algorithm. Learning in sparse coding has been studied extensively in recent years. However, the computation cost for sparse coding is heavy with high dimensional datasets. Four contributions are shown in our work for the problem. First, structure preserving with various dimension reduction methods are studied to improve computational efficiency as well as classification accuracy. Second, Laplacian score dictionary and Dictionary Decision via Active Learning are proposed for dictionary learning. Third, sparse imputation has been used to evaluate the feature for the classification, and selected features are contributed for the classification performance. Fourth, L1 graph is built via sparse coding, and related features are chosen. Extensive empirical studies over synthetic and real-world data have been evaluated for the proposed method, and related theoretic justification has also been conducted.