Prediction of biological data by using weighted ensemble classifiers

Nowadays, cancer disease is rapidly spreading and it is difficult to treat. There are different types of cancer such as breast, lung, central nervous system, leukemia and many more. Each type consists of different sequence error in the gene syntax in the body. In order to treat them, early diagnosis is mandatory, It means that gene sequences must be interpreted in a smart and accurate way. Genome wide scale data classification has been triggered by the microarray technologies. A DNA microarray can have different expression levels of genes simultaneously. Past studies indicate that microarray technology can provide means for tumor classification. Related datasets contain small number of samples; the number of genes is extremely larger than the number of samples and these make knowledge discovery challenging in microarray. For this reason, different classification methods were proposed. However, classifiers' performance results may be dataset dependent. While the classification accuracy of one classifier can make high accuracy for some dataset, it may give poor accuracy for another. That means, there is no perfect classifier works for all datasets robustly. Regardless of selecting one classification approach, ensemble classifiers approach is proposed. It combines each individual classifier prediction in the pool to create joint decision-rules and classify unknown samples according to this decision rules. Ensemble classifiers is being used in many image processing, artificial intelligence and medical fields as well. In this thesis, I worked on creating an ensemble classifiers method that chooses best classifier combination in the classifier pool. I focused on twenty nine different classifiers in literature to selection each having different weights based on their past performance. A two level genetic algorithm is used to create ensemble classifiers. In this proposed method, the first level determines the classifiers to use; the second level optimizes their weights for effectiveness. The results of this study are promising when compared to other methods.