英文摘要 | This paper conducts a preliminary study on Non-Informational Knowledge-based Discriminative Learning. Compared with other related works, our study is distinguished from the following aspects: (1) Unlike purely data-driven learning approaches that targeted at general solutions, we integrate knowledge into learning model to improve performance; (2) Different from ad-hoc approaches that use domain knowledge and are thus problem-dependent, our approaches only take advantage of non-informational knowledge and hence the solutions obtained are totally domain-independent; (3) Compared with the non-informational prior based probabilistic generative learning, all approaches we propose are discriminative. The main content covers four important research topics, i.e., classification, regression, model selection and feature selection, and in particular includes the following: 1. A novel approach to classification, namely the S-learning approach, which has both empirical and theoretical advantages over other counterparts, such as generalization accuracy, computational efficiency, robustness to outliers, theoretical amenability, and natural sparseness in the kernel space; 2. A new regression scheme: reformulated parametric learning, which learns parameters equivalently based on a simpler model in the equivalence set of the original model; 3. Structural identifiability based model selection: a functional framework for parameter dependence examination and to criteria to detect parameter redundancy; 4. A theoretical optimal criteria for feature selection that directly reflects Bayes error, and an algorithmic framework that selects a subset of features by minimizing nonparametric Bayes error of the training data set. We evaluate our proposed algorithms on various benchmark data set and problems in comparison with other state-of-art methods. The experimental results confirm the effectiveness of our proposed approaches. Keywords: Machine Learning, Supervised Learning, Classification, Regression, Model Selection, Qualitative Experiment Design, Dimensionality Reduction, Feature Selection, Loss Function, Kernel Methods, S-Leaning, Reformulated Parametric Learning, Equivalent Decomposition Model, Structural Identifiability, Parameter Redundancy, Parameter Dependence, Nonparametric Bayes Error Minimization, Feature Weighting, Relief |
修改评论