Search

Feature selection for supervised learning concerns the problem of selecting a number of important features (w.r.t. the class labels) for the purposes of training accurate prediction models. Traditional feature selection methods, however, fail to take the sample distributions into consideration which may lead to poor prediction for minority class examples. Due to the sophistication and the cost involved in the data collection process, many applications, such as biomedical research, commonly face biased data collections with one class of examples (e.g., diseased samples) significantly less than other classes (e.g., normal samples). For these applications, the minority class examples, such as disease samples, credit card frauds, and network intrusions, are only a small portion of the data but deserve full attention for accurate prediction. In this paper, we propose three filtering techniques, Higher Weight (HW), Differential Minority Repeat (DMR) and Balanced Minority Repeat (BMR), to identify important features from datasets with biased sample distribution. Experimental comparisons with the ReliefF method on five datasets demonstrate the effectiveness of the proposed methods in selecting informative features for accurate prediction of minority class examples. © 2010 World Scientific Publishing Company.

Feature selection concerns the problem of selecting a number of important features (w.r.t. the class labels) in order to build accurate prediction models. Traditional feature selection methods, however, fail to take the sample distributions into the consideration which may lead to poor predictions for minority class examples. Due to the sophistication and the cost involved in the data collection process, many applications, such as Biomedical research, commonly face biased data collections with one class of examples (e.g., diseased samples) significantly less than other classes (e.g., normal samples). For these applications, the minority class examples, such as disease samples, credit card frauds, and network intrusions, are only a small portion of the data collections but deserve full attentions for accurate prediction. In this paper, we propose three filtering techniques, Higher Weight (HW), Differential Minority Repeat (DMR) and Balanced Minority Repeat (BMR), to identify important features from biased data collections. Experimental comparisons with the ReliefF method on five datasets demonstrate the effectiveness of the proposed methods in selecting informative features from data with biased sample distributions.

Microarray expression data, which contain expression levels of a large number of simultaneously observed genes, have been used in many scientific research and clinical studies. Due to its high dimensionalities, selecting a small number of genes has shown to be beneficial for tasks such as building prediction models for molecular classification of cancers. Traditional gene selection methods, however, fail to take the sample distributions into consideration for gene selection. Due to the scarcity of the samples, in Biomedical research it is very common to have severely biased data distributions with one class of examples (e.g., diseased samples) significantly less than other classes (e.g., normal samples). Sample sets with biased distributions require special attention for identifying genes responsible for particular disease. In this paper, we propose three filtering techniques, Higher Weight (HW), Differential Minority Repeat (DMR) and Balanced Minority Repeat (BMR), to identify genes relevant to fatal diseases for biased microarray expression data. Experimental comparisons with the traditional ReliefF method on five microarray datasets demonstrate the effectiveness of the proposed methods in selecting informative genes from microarray expression data with biased sample distributions.

Microarray experiments usually output small volumes but high dimensional data. Selecting a number of genes relevant to the tasks at hand is usually one of the most important steps for the expression data analysis. While numerous researches have demonstrated the effectiveness of gene selection from different perspectives, existing endeavors, unfortunately, ignore the data imbalance reality, where one type of samples (e.g., cancer tissues) may be significantly fewer than the other (e.g., normal tissues). In this paper, we carry out a systematic study to investigate the impact of gene selection on imbalanced microarray data. Our objective is to understand that if gene selection is applied to imbalanced expression data, what kind of consequences it may bring to the final results? For this purpose, we apply five gene selection measures to eleven microarray datasets, and employ four learning methods to build classification models from the data containing selected genes only. Our study will bring important findings and draw numerous conclusions on (1) the impact of gene selection on imbalanced data, and (2) behaviors of different learning methods on the selected data.

Recent years have seen increasing quantities of high-throughput biological data available for genetic disease profiling, protein structure and function prediction, and new drug and therapy discovery. High-throughput biological experiments output high volume and/or high dimensional data, which impose significant challenges for molecular biologists and domain experts to properly and rapidly digest and interpret the data. In this paper, we provide simple background knowledge for computer scientists to understand how supervised learning tools can be used to solve biological challenges, with a primary focus on two types of problems: Biological sequence profiling and microarray expression data analysis. We employ a set of supervised learning methods to analyze four types of biological data: (1) gene promoter site prediction; (2) splice junction prediction; (3) protein structure prediction; and (4) gene expression data analysis. We argue that although existing studies favor one or two learning methods (such as Support Vector Machines), such conclusions might have been biased, mainly because of the inadequacy of the measures employed in their study. A line of learning algorithms should be considered in different scenarios, depending on the objective and the requirement of the applications, such as the system running time or the prediction accuracy on the minority class examples.