Efficient Splice Site Prediction with Context-Sensitive Distance Kernels

This paper presents a comparison between different context-sensitive kernel functions for doing splice site prediction with a support vector machine. Four types of kernel functions will be used: linear-, polynomial-, radial basis functionand negative distance-based kernels. Domain-knowledge can be incorporated into the kernels by incorporating statistical measures or by directly plugging in distance functions defined on the splice site instances. From the experimental results it becomes clear that the radial basis function-based kernels get the best accuracies. However, because classification speed is of crucial importance to the splice site prediction system, this kernel is computationally too expensive. Nevertheless, in general incorporating domain knowledge does not only improve classification accuracy, but also reduces model complexity which in its turn again increases classification speed.