A Neural Framework for Web Ranking Using Combination of Content and Context Features

Containing enormous amounts of various types of data, web has become the main source for finding the desired information. Meanwhile retrieving the desired information in such a vast heterogeneous environment is much difficult. This situation has led to a drastic increase in the popularity of internet search engines. Undoubtedly, designing both efficient and effective ranking strategies as the basic core of web information retrieval systems are unavoidable. Unfortunately most of the proposed ranking algorithms do not work very well over general datasets because of their fixed configurations. Many of these algorithms also suffer from their computational costs. Regarding these shortcoming, in this paper, a new ranking framework named "NNRank" is proposed which uses the primitive features of web documents from the categories of content and context using an artificial neural network. The neural networks selected in our approach is a radial basis function or a principle component analysis neural network which due to their high convergence rate, have the capability to exhibit a high performance with a limited number of features. Experimental results based on TREC 2004 gathered in Microsoft LETOR dataset, indicate a noticeable enhancement comparing to the well-known ranking algorithms such as TF-IDF, PageRank and HITS. The results are also comparable with those of BM25.