Let terms choose their own kernels: An intelligent approach to kernel selection for healthcare search

Proximity-based information retrieval models usually pre-define a kernel function on all documents. However, terms, especially professional terms used in healthcare, can exert various levels of influence in different positions. In this paper, we propose an intelligent approach with an elite kernel function for healthcare search. The unique contribution is that the proposed approach assigns a term with the most suitable kernel function at a position, instead of pre-defining a single kernel, such as the Gaussian kernel, over all documents. Furthermore, our approach requires no prior data, external resources, or advance training steps. We first define the elite kernel and present its existence in a kernel space. Then, we optimize the kernel by maximizing the probability using the Poisson process. Finally, we use a kernel-based weighting function to evaluate the effectiveness of the intelligent approach on three standard CLEF healthcare domain data sets and two non-healthcare domain data sets. Our approach shows great advantages in terms of both healthcare domain and non-healthcare domain data.