A Belief Theory Based Instance Selection Scheme for Label Noise and Outlier Detection from Breast Cancer Data

In case of real datasets, the likelihood of the training data being corrupted with training label noise and outliers arises. Certain classification algorithms including support vector machine (SVM) is sensitive to noise and outlier samples which can degrade their performance. Belief theory which involves an extension of the general probabilistic model and utilises combination rules for information fusion has found good use in the realm of classifiers. In this paper, we propose a belief theory based instance selection (BIS) scheme using the k nearest neighbours (KNN) algorithm for removing outlier and noise samples prior to SVM training to increase classification performance for breast cancer FNAC (Fine needle aspiration cytology) image data features. Our algorithm is tested on the WBCD database from the UCI machine learning repository which contains FNAC image data features. Performance evaluation is done by considering accuracy and confusion matrix measures. Effect of noise is assessed by testing on the datasets after contaminating the training data by random mislabelling. Results are compared with the conventional SVM algorithm for both the noisy and noiseless datasets. The proposed BIS scheme is shown to improve the performance of the SVM classifier considerably under noisy conditions.