A0671
Title: Transformation and covariance estimation for the non-linearly separable misclassification problem
Authors: Mubarak AL-Shukeili - Sultan Qaboos University (Oman) [presenting]
Ronald Wesonga - Sultan Qaboos University (Oman)
Abstract: The search for a suitable binary classifier mainly depends on the location vector and covariance matrix. The linear discriminant, for example, works by constructing the most suitable linear hyperplane based on location and covariance among parameters. The data point $X$ are transformed from $R^p$ to $R$ such that the resultant data leads to a minimum misclassification rate. However, when the class data are significantly overlapped due to class homogeneity, linear classifiers perform poorly. We present a method that performs the transformation of classes to be linearly separable prior to classification. Moreover, our study also proposed an estimate of a covariance matrix for one class given the other class is known. The resultant performance of the proposed method is validated using simulation and real-life data. Findings show that our method yields more competitive results compared to the classical quadratic discriminant analysis
