Restricting Supervised Learning: Feature Selection and Feature Space Partition

Many supervised learning problems are considered difficult to solve either because of the redundant features or because of the structural complexity of the generative function. Redundant features increase the learning noise and therefore decrease the prediction performance. Additionally, a number of problems in various applications such as bioinformatics or image processing, whose data are sampled in a high dimensional space, suffer the curse of dimensionality, and there are not enough observations to obtain good estimates. Therefore, it is necessary to reduce such features under consideration. Another issue of supervised learning is caused by the complexity of an unknown generative model. To obtain a low variance predictor, linear or other simple functions are normally suggested, but they usually result in high bias. Hence, a possible solution is to partition the feature space into multiple non-overlapping regions such that each region is simple enough to be classified easily. In this dissertation, we proposed several novel techniques for restricting supervised learning problems with respect to either feature selection or feature space partition. Among different feature selection methods, 1-norm regularization is advocated by many researchers because it incorporates feature selection as part of the learning process. We give special focus here on ranking problems because very little work has been done for ranking using L1 penalty. We present here a 1-norm support vector machine method to simultaneously find a linear ranking function and to perform feature subset selection in ranking problems. Additionally, because ranking is formulated as a classification task when pair-wise data are considered, it increases the computational complexity from linear to quadratic in terms of sample size. We also propose a convex hull reduction method to reduce this impact. The method was tested on one artificial data set and two benchmark real data sets, concrete compressive strength set and Abalone data set. Theoretically, by tuning the trade-off parameter between the 1-norm penalty and the empirical error, any desired size of feature subset could be achieved, but computing the whole solution path in terms of the trade-off parameter is extremely difficult. Therefore, using 1-norm regularization alone may not end up with a feature subset of small size. We propose a recursive feature selection method based on 1-norm regularization which can handle the multi-class setting effectively and efficiently. The selection is performed iteratively. In each iteration, a linear multi-class classifier is trained using 1-norm regularization, which leads to sparse weight vectors, i.e., many feature weights are exactly zero. Those zero-weight features are eliminated in the next iteration. The selection process has a fast rate of convergence. We tested our method on an earthworm microarray data set and the empirical results demonstrate that the selected features (genes) have very competitive discriminative power. Feature space partition separates a complex learning problem into multiple non-overlapping simple sub-problems. It is normally implemented in a hierarchical fashion. Different from decision tree, a leaf node of this hierarchical structure does not represent a single decision, but represents a region (sub-problem) that is solvable with respect to linear functions or other simple functions. In our work, we incorporate domain knowledge in the feature space partition process. We consider domain information encoded by discrete or categorical attributes. A discrete or categorical attribute provides a natural partition of the problem domain, and hence divides the original problem into several non-overlapping sub-problems. In this sense, the domain information is useful if the partition simplifies the learning task. However it is not trivial to select the discrete or categorical attribute that maximally simplify the learning task. A naive approach exhaustively searches all the possible restructured problems. It is computationally prohibitive when the number of discrete or categorical attributes is large. We describe a metric to rank attributes according to their potential to reduce the uncertainty of a classification task. It is quantified as a conditional entropy achieved using a set of optimal classifiers, each of which is built for a sub-problem defined by the attribute under consideration. To avoid high computational cost, we approximate the solution by the expected minimum conditional entropy with respect to random projections. This approach was tested on three artificial data sets, three cheminformatics data sets, and two leukemia gene expression data sets. Empirical results demonstrate that our method is capable of selecting a proper discrete or categorical attribute to simplify the problem, i.e., the performance of the classifier built for the restructured problem always beats that of the original problem. Restricting supervised learning is always about building simple learning functions using a limited number of features. Top Selected Pair (TSP) method builds simple classifiers based on very few (for example, two) features with simple arithmetic calculation. However, traditional TSP method only deals with static data. In this dissertation, we propose classification methods for time series data that only depend on a few pairs of features. Based on the different comparison strategies, we developed the following approaches: TSP based on average, TSP based on trend, and TSP based on trend and absolute difference amount. In addition, inspired by the idea of using two features, we propose a time series classification method based on few feature pairs using dynamic time warping and nearest neighbor