A Study Of GPU-based Parallel Learning To Rank

The emergence of search engines helps people find relevant information in theInternet. Sorting of retrieved results is crucial. Learning to rank, as a novelinformation retrieval technique, is a new solution to the problem of informationretrieval. The traditional research of learning to rank algorithm is based onsmall-scale datasets. Due to the scale of Internet information that increases rapidly,it is challenging for many existing algorithms to handle such large-scale data. Theperformance of learning to rank has become a serious problem. This paper presents anew learning to rank algorithm and combines graphics processing unit (GPU)parallel computing technology and verifies the performance of the algorithm. Thecontent of the paper contains several points as follows.（1） Summarize and elaborate the theory of learning to rank and GPU parallelcomputing. Describe the existing learning to rank algorithms, expounded evaluationcriteria and parallel programming model.（2） In-depth analysis of information retrieval technology and combined withthe combination of the characteristics of the more relevant information are veryimportant. This paper adopts pair-wise learning to rank research direction. Weredivided the data input space and used larger partial order document pairs as theinput space.（3）This paper proposes a learning to rank algorithm based on Bayesianpersonal ranking methodology. That is a linearly scoring learning to rank model,LSLRM. Solve the ranking problem by estimating the right rank on the inputdocument pairs to build the learning model. That is converted to binaryclassification problems. And then we find the dominant features for distinguishingthe relevance.（4）The algorithm combines the GPU parallel programming model andmemory model to solve the performance bottlenecks of learning to rank inlarge-scale data.（5）Experimental results show the promising performance of GPU-basedparallel learning to rank. We compared our algorithm with RankSVM-Struct, and soon, benchmark algorithm on benchmark data sets released by Microsoft ResearchAsia. The results show that our proposed algorithm superior to other algorithms. Thetime performance of our algorithm can be10-11times faster than that on CPU.