Kailash Shaw

An enhanced classifier fusion model for classifying biomedical data

Classification is a technique where we discover the hidden class level of the unknown data. As different classification methods produces different accuracy according to the class level; classifier fusion is the solution to achieve more accuracy in every level of the input data. Selection of a suitable classifier in classifier fusion is a tedious task. In the proposed model, the output of the three classifiers is fed to the dynamic classifier fusion technique. This model will use each classifier for every individual data. We have used principal component analysis (PCA) to deal with issues of high dimensionality in biomedical classification. Three types of classification techniques on microarray data like multi layer perceptron (MLP), FLANN and PSO-FLANN have been implemented and compared; it has been observed that MLP is showing better result. We have also proposed a model for classifier fusion, where the model will choose the relevant classifiers according to the different region of datasets.

Gene expression network discovery: a pattern based biclustering approach

Discovering biologically significant information from gene expression data is now a days playing important role in gene function detection, gene regulation, drug discovery, detecting and predicting the diseases. Many traditional clustering algorithms are present to discover such gene regulations. Such discovered clusters are known as global clusters, which incurs more processing overhead. To overcome such problem, the biclustering approach, also known as local clustering has been emerged. Generally, there are two ways of measuring the similarity among subset of objects and attributes. First one is grouping the data elements by measuring the similarity based on the proximity. But, there may be the case that, many objects and attributes which are far apart but the gives significant meaning for being grouped. This problem can be solved by the second method, which not only measures the proximity of data elements but also find subset of objects and attributes which forms similar or coherent patterns such as scaling and shifting irrespective of their proximity. In this paper, we have implemented the pattern based clustering and before that the dimensionality reduction using Principal Component Analysis (PCA) is used to reduce the attributes without loss of information. We have compared the Minimum Squared Residue (MSR) approach of Cheng and Church with our proposed model. Our method shows its better performance as compared to MSR based approach.

Discovering non-exclusive functional modules from gene expression data

Biological processes are not independent of each other as genes participate in multiple different processes. Each gene should be assigned to multiple biclusters. In real life, more than one gene is responsible for a particular type of disease. The biclustering can associate clusters with gene arrangement patterns, preserving genomic information. Additionally, overlapping capability is desirable for the discovery of multiple conserved patterns within a single genome. In strict or crisp partition-based biclustering, each gene/condition belongs to exactly one functional module whereas, addressing some biological questions requires partitioning methods leading to non-exclusive functional modules. The proposed method involves a novel strategy to discover such non-exclusive pattern-based biclusters using fuzzy set approach. We have evaluated the performance of our proposed model with few existing ones and the result shows that this can be suitable for application to genomes with high genetic exchange and various conserved gene arrangements in gene regulatory networks.

Clustering and classifying informative attributes using rough set theory

Clustering techniques are the unsupervised data mining applications and are important in data mining methods for exploring natural structure and identifying interesting patterns in original data, also it is proved to be helpful in finding coexpressed samples. In cluster analysis, generally the given dataset is partitioned into groups based on the given features such that the data objects in the same group are more similar to each other than the data objects in other groups. The objects are clustered or grouped based on the principle of maximizing intra-class similarity and minimizing interclass similarity. In this paper, the rough set theory (RST) has been used for attribute clustering. RST is a theory adopted to deal with rough and unsure knowledge, which analyzes the clusters and finds the data principles when previous knowledge is not available, providing a new method for data classification. With the continuous change in data objects we have to improve these relevant technologies over time, and we have to propose creative theory in response, meeting the demands of application, though there are many rough set methods. In this paper; after implementing the rough set based attribute clustering method on real life leukemia dataset, we classify them using some of the traditional classification techniques such as Multilayered Perceptron (MLP) based classifier, Naïve Bayesian (NB) classifier and Support Vector Machine (SVM). At the end, the same classification techniques are applied to classify the original leukemia dataset before application of rough set based attribute clustering. Finally the paper provides a comparative analysis among the traditional classifiers and the proposed corresponding rough set based classifiers. Among all, the proposed MLP classifier is found to be the better classifier than the others giving higher classification accuracy and it is proved to be efficient having lower error ratio.

Impact of learning algorithms on gene expression data set

Classification is a process which plays a vital role in the analysis of the gene expression data set. The paper focuses on variety of learning algorithms which are really challenging in nature. The proposed model has been implemented and evaluated by using 5 benchmark datasets and to evaluate the performance and throughput of the model, various learning algorithms has been used like Random Forest, Support vector Machine, K-Nearest Neighbor, Bayesian, Linear Discriminate, Multi layer Perception and Decision Tree. We proposed model by using the k –fold cross validation for training and testing of the data.

A semi-supervised rough set and random forest approach for pattern classification of gene expression data

In this paper, we present a semi-supervised rough set-based random forest gene selection method for classification of data patterns. The proposed method tries to find the genes of interest known as significant genes and maximise the accuracy of the model with reduction percentage. The advantage of this approach is analysed by experimental results on three benchmark datasets such as leukaemia, colon cancer and SRBCT and results showed an improved accuracy over existing methods such as support vector machine, k-nearest neighbour and random forest. Finally, the performance of those selected significant genes has been measured using classifier validity and statistical measures. The experimental results and performance measures proves the efficiency of the proposed hybridised technique over traditional random forest method.

Hash based biclustering for class discovery from gene expression data: A pattern similarity approach

Cellular processes come forth on subsets of genes to be co-expressed and correlated under certain experimental conditions, but behaves almost independently under other conditions. So, discovering such local expression patterns may uncover many genetic mechanisms which lead to class discovery. In this paper, we have proposed an efficient hash based biclustering approach, which identifies coherent patterns known as scaling and shifting patterns from high dimensional gene expression datasets. Our proposed algorithm consists of two steps: first, we pre-process the data set by reducing the attributes without much loss of information using Principal Component Analysis (PCA) and second, an enhanced pCluster algorithm using hashing technique to make the searching faster is used to discover the scaling and shifting patterns, which leads to patter based biclusters and those biclusters will contribute for class discovery. Finally, we have compared our method with some existing pattern based models and it has been found that our algorithm is very versatile and promising.