© 2018 IEEE. Classification of data containing disproportionate class distributions or rare events, proves troublesome for groups of models. Kernel Logistic Regression (KLR) is such a framework which produces a model biased in favour of the majority class, when classes are severely imbalanced. A weighted form of the likelihood function has been proposed, which successfully improves model performance on the minority class. This modification of the likelihood, which is dependent on the number of instances from the minority class contained in the training set however, does not account for the possibility of incorrect training labels. Consequently, a correction to the expression of the weighted likelihood function for KLR, which is robust to the presence of noise in the training label set, is proposed in this paper. The resultant model, allows for superior performance in a noisy and noise free setting to be achieved. The performance is compared against the original likelihood and weighted likelihood functions on several benchmark examples.