Jianguo Lee

Classification of gene-expression data: The manifold-based metric learning way

Classification of microarray gene-expression data can potentially help medical diagnosis, and becomes an important topic in bioinformatics. However, microarray data sets are usually of small sample size relative to an overwhelming number of genes. This makes the classification problem fairly challenging. Instance-based learning (IBL) algorithms, such as nearest neighbor (k-NN), are usually the baseline algorithm due to their simplicity. However, practices show that k-NN performs not very well in this field. This paper introduces manifold-based metric learning to improve the performance of IBL methods. A novel metric learning algorithm is proposed by utilizing both local manifold structural information and local discriminant information. In addition, a random subspace extension is also presented. We apply the proposed algorithm to the gene-classification problem in three ways: one in the original feature space, another in the reduced feature space, and the third via the random subspace extension. Statistical evaluation shows that the proposed algorithm can achieve promising results, and gain significant performance improvement over traditional IBL algorithms.

Kernel trick embedded Gaussian mixture model

In this paper, we present a kernel trick embedded Gaussian Mixture Model (GMM), called kernel GMM. The basic idea is to embed kernel trick into EM algorithm and deduce a parameter estimation algorithm for GMM in feature space. Kernel GMM could be viewed as a Bayesian Kernel Method. Compared with most classical kernel methods, the proposed method can solve problems in probabilistic framework. Moreover, it can tackle nonlinear problems better than the traditional GMM. To avoid great computational cost problem existing in most kernel methods upon large scale data set, we also employ a Monte Carlo sampling technique to speed up kernel GMM so that it is more practical and efficient. Experimental results on synthetic and real-world data set demonstrate that the proposed approach has satisfing performance.

Semi-supervised discriminative classification with application to tumorous tissues segmentation of MR brain images

Due to the large data size of 3D MR brain images and the blurry boundary of the pathological tissues, tumor segmentation work is difficult. This paper introduces a discriminative classification algorithm for semi-automated segmentation of brain tumorous tissues. The classifier uses interactive hints to obtain models to classify normal and tumor tissues. A non-parametric Bayesian Gaussian random field in the semi-supervised mode is implemented. Our approach uses both labeled data and a subset of unlabeled data sampling from 2D/3D images for training the model. Fast algorithm is also developed. Experiments show that our approach produces satisfactory segmentation results comparing to the manually labeled results by experts.

Probabilistic tangent subspace: a unified view

Tangent Distance (TD) is one classical method for invariant pattern classification. However, conventional TD need pre-obtain tangent vectors, which is difficult except for image objects. This paper extends TD to more general pattern classification tasks. The basic assumption is that tangent vectors can be approximately represented by the pattern variations. We propose three probabilistic subspace models to encode the variations: the linear subspace, nonlinear subspace, and manifold subspace models. These three models are addressed in a unified view, namely Probabilistic Tangent Subspace (PTS). Experiments show that PTS can achieve promising classification performance in non-image data sets.

A Discriminative Method For Semi-Automated Tumorous Tissues Segmentation of MR Brain Images

This paper introduces a discriminative method for semiautomated segmentation of the tumorous tissues. Due to the large data of 3D MR brain images and the blurry boundary of the pathological tissues, the segmentation is difficult. A non-parametric Bayesian Gaussian process is proposed to be used for the semi-supervised mode. This discriminative method uses both labeled data and a subset of unlabeled data sampling from 2D/3D images to classify the remains, which is called inductive problem. We propose the prior of traditional Gaussian process to be based on graph regularization and develop a new conditional probability named Extended Bernoulli Model to realize the induction. Fast algorithm to speed up the training phase is also implemented. Experimental results show our approach produces satisfactory segmentations corresponding to the manually labeled results by experts.

Visual object recognition using probabilistic kernel subspace similarity

Probabilistic subspace similarity-based face matching is an efficient face recognition algorithm proposed by Moghaddam et al. It makes one basic assumption: the intra-class face image set spans a linear space. However, there are yet no rational geometric interpretations of the similarity under that assumption. This paper investigates two subjects. First, we present one interpretation of the intra-class linear subspace assumption from the perspective of manifold analysis, and thus discover the geometric nature of the similarity. Second, we also note that the linear subspace assumption does not hold in some cases, and generalize it to nonlinear cases by introducing kernel tricks. The proposed model is named probabilistic kernel subspace similarity (PKSS). Experiments on synthetic data and real visual object recognition tasks show that PKSS can achieve promising performance, and outperform many other current popular object recognition algorithms.

Graph based multi-class semi-supervised learning using gaussian process

This paper proposes a multi-class semi-supervised learning algorithm of the graph based method. We make use of the Bayesian framework of Gaussian process to solve this problem. We propose the prior based on the normalized graph Laplacian, and introduce a new likelihood based on softmax function model. Both the transductive and inductive problems are regarded as MAP (Maximum A Posterior) problems. Experimental results show that our method is competitive with the existing semi-supervised transductive and inductive methods.

Learning probabilistic kernel feature subspace with side-information for classification

Kernel PCA is an efficient method for nonlinear feature extraction. We address two issues in kernel PCA based feature extraction and classification. First, it extracts features without utilizing sample label information. Second, it does not provide a practical means to choose the dimensionality for principal subspace. In this paper, one kind of side-information is incorporated into kernel PCA to solve the first problem. And a complete probabilistic density function is estimated in kernel space so that the choice of dimensionality for principal subspace becomes less important. The proposed model is named probabilistic kernel feature subspace (PKFS). Experiments show that it achieves promising performance and outperforms many other algorithms in classification.