Göksel Biricik

A new method for attribute extraction with application on text classification

We introduce a new method for dimensionality reduction by attribute extraction and evaluate its impact on text classification. The textual contents in body sections of the news in Reuters-21758 are the selected attributes for classification. Using the offered method, high dimension of attributes- words extracted from the news bodies- are projected onto a new hyper plane having dimensions equal to the number of classes. Results show that processing times of classification algorithms dramatically decrease with the attribute extraction method we offer. This is achieved by the fall of the number of attributes given to classifiers. Accuracies of the classification algorithms also increase compared to tests run without using the proposed method.

ARNetMiT R Package: association rules based gene co-expression networks of miRNA targets

miRNAs are key regulators that bind to target genes to suppress their gene expression level. The relations between miRNA-target genes enable users to derive co-expressed genes that may be involved in similar biological processes and functions in cells. We hypothesize that target genes of miRNAs are co-expressed, when they are regulated by multiple miRNAs. With the usage of these co-expressed genes, we can theoretically construct co-expression networks (GCNs) related to 152 diseases. In this study, we introduce ARNetMiT that utilize a hash based association rule algorithm in a novel way to infer the GCNs on miRNA-target genes data. We also present R package of ARNetMiT, which infers and visualizes GCNs of diseases that are selected by users. Our approach assumes miRNAs as transactions and target genes as their items. Support and confidence values are used to prune association rules on miRNA-target genes data to construct support based GCNs (sGCNs) along with support and confidence based GCNs (scGCNs). We use overlap analysis and the topological features for the performance analysis of GCNs. We also infer GCNs with popular GNI algorithms for comparison with the GCNs of ARNetMiT. Overlap analysis results show that ARNetMiT outperforms the compared GNI algorithms. We see that using high confidence values in scGCNs increase the ratio of the overlapped gene-gene interactions between the compared methods. According to the evaluation of the topological features of ARNetMiT based GCNs, the degrees of nodes have power-law distribution. The hub genes discovered by ARNetMiT based GCNs are consistent with the literature.

Solving test suite reduction problem using greedy and genetic algorithms

Regression testing is an important process for software quality. Test case reduction is one of the widely used techniques for regression testing, which can dramatically decrease the testing costs. However, it is an NP-complete problem and big test cases cannot be accomplished in reasonable amount of time. For this reason, we propose a test suite reduction approach by using greedy and genetic algorithms. The greedy algorithm found a wide usage in previous studies thanks to its simplicity, but we already know that it sticks to local optima and does not benefit from metaheuristics. Thus, we used genetic algorithm to overcome its weaknesses. Our experimental results prove that metaheuristics and evolutionary algorithms perform better than the greedy approaches.

A combinatorial approach to construct core and generic gene co-expression networks of colon cancer

Biological experiments can be set in order to detect the causes of diseases. However, they are expensive and time consuming. Recent developments in sequencing technologies help researchers to more easily reveal the underlying mechanisms of the diseases. In this study, we propose a combinatorial method to construct generic and core gene co-expression networks (GCNs) to discover the genes and their interactions related to colon cancer. We apply five gene network inference (GNI) algorithms and combine their estimations with Simple Majority Voting to specify the frequently inferred gene interactions and obtain the resulting GCNs on two different gene expression datasets. We then apply the intersection and union operators on these GCNS to derive the core and generic GCNs, respectively. The evaluation results of overlap analysis and topological features of GCNs for the colon cancer show that the networks produced with the proposed approach fit to the power-law degree distribution better.

Comparing the impacts of dimension reduction methods that use class labels on text classification

Classification of datasets that contain samples with numerous features is known as a costly process in time and space. In order to overcome this problem, dimensionality reduction techniques like feature selection and feature extraction are proposed in literature. In this paper, we compare the impacts of abstract feature extraction method and other popular techniques that use class labels for dimensionality reduction on classification performances. For evaluation, we utilize two standard text datasets having high dimensional samples. We compare the impacts of selected methods on performance by applying them on selected datasets and testing on five different classifiers with different design approaches. Results show that using abstract feature extraction method for dimensionality reduction produces much better classification performance, when compared with other selected methods.

Demographic information classification exploiting spoken language

Recently, extracting the demographic information like age, gender and race by using speech and face attributes takes much attention in the literature. In this research, we have focused on the implementation of a demographic information classification system and proved the relationship between spoken language and demographic profile of people. In the first step, the feature vectors of spoken language were extracted then dimensions of the feature vectors were reduced by our feature reduction method and Correlation Based Feature Selection method. Finally, the success of Naïve Bayes, Support Vector Machine and K-Nearest Neigbour classification algorithms was evaluated.