Guihua Wen

Enhanced supervised locality preserving projections for face recognition

To address the problem of “curse of dimensionality”, usually dimensionality reduction is used to reduce datas dimensionalities. As a graph-based method for linear dimensionality reduction, Locality Preserving Projections (LPP) searches for an embedding space in which the similarity among the local neighborhoods is preserved. However, LPP has two disadvantages: Firstly, LPP doesnt take the label information into consideration which is crucial for classification tasks; Secondly, like most graph-based methods, graph construction of LPP is sensitive to noise and outliers. To these end, we propose an Enhanced Supervised LPP(ESLPP) that allows both locality and class label information to be incorporated which improves the performance of classification. In the mean time, ESLPP uses similarity based on robust path instead of Gaussian heat kernel similarity such that it can capture the underlying geometric distribution of samples even when there are noise and outliers. Experimental results on face databases confirm its effectiveness.

Cognitive facial expression recognition with constrained dimensionality reduction

Facial expression recognition (FER) is an important research area in human-computer interaction. In this paper, a new dimensionality reduction method together with a new classifier are proposed for FER. The goals of most dimensionality reduction contains minimizing the within-class distances. However, the within-class distances for some expressions could be very large, so that to minimize these distances could largely influence the optimization function. To overcome this defect, a new dimensionality reduction method is proposed by adding a penalty item, which is the sum of within distances that are far from each other. Through maximizing this item, the distances among faces with the same expression that are far from each other cannot be minimized to too small. Besides, this method can partly characterize the density information from training samples. To make full use of density information, a new classification method is developed that is based on the enhanced cognitive gravity model. The conducted experiments validate the proposed approach in term of the performance of facial expression recognition. The approach presents the excellent performance over previously available techniques.

Neural gas based cluster ensemble algorithm and its application to cancer data

The cluster ensemble approach is gaining more and more attention in recent years due to its useful applications in bioinformatics and pattern recognition. In this paper, we present a new cluster ensemble approach named as the neural gas based cluster ensemble algorithm (NGCEA) for class discovery from biological meaningful data, NGCEA first adopts the perturbed function to generate a set of new datasets. Then, it proposes to adopt the neural gas algorithm to obtain the clustering solutions from the perturbed datasets, In the following, NGCEA views the row of each clustering solution as the new features, and forms a new dataset. Finally, it adopts the neural gas algorithm as consensus function to perform clustering again on the new dataset and obtains the final result. The experiments in cancer datasets show that (i)NGCEA works well on most of cancer datasets (ii) NGCEA outperforms most of the state-of-the-art cluster ensemble algorithms when applied to gene expression data

Audio feature extraction for classification using relative transformation

Audio feature extraction plays a much important role in the areas of audio processing. This paper proposes a new audio feature extraction method using the relative transformation (RT). It begins with equally dividing an audio signal into a lot of segments. On each segment, the mel-frequency cepstral coefficients are extracted and combined by RT to generate a single feature vector. All these vectors are then combined by RT again to generate a single feature vector for the audio. This method can nicely deal with the noisy, sparse, and imbalance problems, while it has lower time complexity. It is purely data-driven and does not depend on particular audio characteristics. The experimental results suggest that the classifier with the proposed method often gives the better results in classification.

Erratum to: Relative manifold based semi-supervised dimensionality reduction

Figure 8 of this article shows YaleB and CMU PIE with incorrect legend titles: YaleB (Tr=1900, Te=514, NOC=100) should be YaleB (Tr=1900, Te=514, d=100) (Fig. 8(a)); TIE (Tr=1200, Te=2880, d=100) should be PIE (Tr=1200, Te=2880, d=100) (Fig. 8(b)).In Fig. 9, the legend keys and the legend texts are mismatched. The correct figure is illustrated as follows.

Penalty-based cluster validity index for class discovery from cancer data

In order to perform successful diagnosis and treatment of cancer, discovering and classifying cancer types correctly is essential. One of the challenges in cancer class discovery is to estimate the number of classes given a set of unknown microarray data. In the paper, we propose a new cluster validity criterion called Penalty-based Disagreement Index (PDI) based on the perturbation technique to estimate the number of classes in microarray data, PDI not only considers the disagreement between the partition results obtained from the original data and those obtained from the perturbed data, but also includes a penalty measure which is a function of the number of classes. Our experiments show that PDI successfully estimates the true number of classes in a number of challenging real cancer datasets.

PSEFminer: A new probabilistic subspace ensemble framework for cancer microarray data analysis

In order to perform successful diagnosis and treatment of cancer, discovering and classifying cancer types correctly is essential. Most of the existing works adopt single clustering algorithms to perform class discovery from bio-molecular data. Unfortunately, single clustering algorithms have limitations, which are lack of the robustness, stableness and accuracy. In this paper, we develop a new probabilistic subspace ensemble framework known as PSEFminer for cancer microarray data analysis. PSEFminer integrates the probabilistic subspace generator, the self-organizing map(SOM) and the normalized cut algorithm into the ensemble framework to discover the underlying structure from cancer microarray data. The experiments in cancer datasets show that (i) the probabilistic subspace generator plays an important role to improve the performance of PSEFminer; (ii) PSEFminer outperforms most of the state-of-the-art cluster ensemble algorithms when applied to cancer gene expression data

Relative nearest neighbors for classification

Classification approaches based on the k nearest neighbors are simple and often result in goodxii performance. However, they heavily depend on the collection of selected neighbors. When performing the classification on the sparse or noisy data, the selected nearest neighbors are not consistent with our perception which in turn leads to the worse performance. This paper proposes two new classifiers by applying the relative transformation to define the k nearest neighbors, where the relative transformation is defined on the local region varying with the query sample to generate the relative space in which nearest neighbors to the query sample can be selected more reason-ably. The conducted experiments on challenging benchmark data sets validate the proposed approach.

A Neighborhood Preserving Based Semi-supervised Dimensionality Reduction Method for Cancer Classification

Cancer classification of gene expression data helps determine appropriate treatment and the prognosis. Accurate prediction to the type or size of tumors relies on adopting efficient classification models such that patients can be provided with better treatment to therapy. In order to gain better classification, in this study, a linear relevant feature dimensionality reduction method termed the neighborhood preserving based semi-supervised dimensionality reduction (NPSSDR) is applied. Different from traditional supervised or unsupervised methods, NPSSDR makes full use of side information, which not only preserves the must-link and cannot-link constraints but also can preserve the local structure of the input data in the low dimensional embedding subspace. Experimental results using public gene expression data show the superior performance of the method.

Improved locally linear embedding by cognitive geometry

Locally linear embedding heavily depends on whether the neighborhood graph represents the underlying geometry structure of the data manifolds. Inspired from the cognitive relativity, this paper proposes a relative transformation that can be applied to build the relative space from the original space of data. In relative space, the noise and outliers will become further away from the normal points, while the near points will become relative closer. Accordingly we determine the neighborhood in the relative space for Hessian locally linear embedding, while the embedding is still performed in the original space. The conducted experiments on both synthetic and real data sets validate the approach.