A Ranking Algorithm ListNet Based On Stochastic Gradient Descent

With the rapid development of the Internet, search engine is more and more important. It's very important that how to effectively search useful information, a good search engine can help users save a lot of time. Search engine's core part is how to rank web page associated to inquires. Page ranking's stand or fall is becoming more and more important, which decides whether a search engine is good. Learning to rank is the current popular page ranking field, and has good performance. According to the ranking method, it can be divided into three categories: pointwise, pairwise and listwise. And listwise approach usually performs better than the other approaches. ListNet is a classical and effect listwise approach. But also has shortcomings: ListNet algorithm convergence slowly, and its running time directly depends on the size of training set and is not suited to train a data set with large quantity queries , not to mention mass datasets.Nowadays, as there are more and more large data sets, and even mass datasets, existing algorithms of learning to rank have not been suited to train a data set with large quantity queries , not to mention mass datasets. Because the existing algorithms need to read the training data into memory for one time, but large data sets, and even mass datasets is not possible to be read into memory once. Therefore, the motivation of this paper is to design a algorithm of learning to rank that can be applied to train the large datasets, and even mass datasets.This paper proposes the improvement ideas of ListNet algorithm. The original algorithm use neural network as model, and now we use SVM. We use Stochastic Gradient Descent algorithm instead of the original gradient descent algorithm. We combine ListNet and Pegasos to make up the shortfall, make the improved algorithm convergence faster, make the running time does not directly depends on the size of the training set and make the improved algorithm can be suited to train a data set with large quantity queries, even mass datasets.In this paper, we use LETOR MQ2007,MQ2008, OHSUMED TD2003, TD2004 data sets to do the experiments. First, we compare running time of the improved algorithm and ListNet's. Second, we compare the accuracy of the improved algorithm and ListNet's. Last, we compare the accuracy of the improved algorithm and other listwise approaches. Experimental results show that relative to ListNet, the improved algorithm has absolute advantage in the training time, and the improved algorithm performs better than ListNet and other listwise approaches in data set with large quantity queries.