Research On Expert List Sorting Learning Method Based On Sparse Learning

Expert search are currently hot field in information retrieval research field. Currently, the general expert retrieval systems are the way to sort to return to the user the query, so the core issues expert retrieval can be transformed into how efficiently sorted. Sort by today’s mainstream method is to use machine learning to sort, divided based on the data pointwise, based on data pairwise and sorting method based on data listwise. This article is from the perspective of an expert on the list-based sorting, sort a list of experts for the presence of traditional training time is too long, easy to fall into local optimal solution, the paper considers the use of sparse algorithm combining expert list sort to build a model to solve the above problem. Mainly in the following aspects of the exploration and research, and have achieved some results:(1) according to the characteristics of the expert sorting task, studies the influence factors affect experts sorting, defines the characteristics of four categories, including query-document similarity features, language model, content characteristic and the experts related features. After a follow-up experiments show that in the above four categories characteristics can effectively improve the accuracy of expert sort.(2) the traditonal learning rank presence the problems of random initialization parameters easily understood cause local optima, a higher dimension characteristics cause training time is too long, and Sparse algorithm for high-dimensional data can be linearly decomposed and solved, effectively shortening the training time. Therefore, the sort learning method proposed based on sparse learning expert list. The method from the perspective of direct optimization of IR measuring performance indicators, first define the expert list loss function, and then according to defined target 1-norm thinning algorithm optimization function, then the model parameters β tuning in the objective function, to get the global optimal solution, finally, ordered by β descending order to realize the experts, the subsequent experiments show that the fusion sparse learning can effectively improve the effect of expert sort.(3) based on the higher dimension expert characteristics redundancy to reduce the sorting accuracy easily, optimization of characteristics of experts, and then to carry on the model parameter tuning, was proposed based on sparse learning of two-stage expert list ordering method. This method first define the objective function by the sparse algorithm combining with expert list sorting, and using the theory of sparse threshold value calculating characteristics, characteristics of the optimization filter, finally, the characteristics of the threshold and the model parameter vector β in the objective function are applied to solve the cross validation, finding the optimal model parameter vector β and ordered by β descending order to experts, the effectiveness of the proposed method has been proved by the experiment.