Enhancing the learning to rank using the virtual feature logistic regression with relevance feedback

Many information retrieval applications have to publish their outputs in the form of ranked lists, in which documents must be sorted in descending order according to their relevance to a given query. Many existing methods perform analysis on multidimensional features distilled from query-document pairs directly and don´t take user´s interactive feedback into account; hence, they incur a high computation overhead and a low retrieval performance due to inaccurate query expression. In this paper, we propose a Virtual Feature Logistic Regression (VFLR) method that conducts the logistic regression on a set of crucial but independent variables, called virtual features (VF), which are extracted by the principal component analysis (PCA) with the user´s relevance feedback. We then predict the ranking score of each queried document to produce a ranked list. We systematically evaluate our method using the MQ2008 dataset. The experimental results validate that the VFLR method outperforms the state-of-the-art methods in terms of the Mean Average Precision (MAP), the Precision at position k (P@k), and the Normalized Discounted Cumulative Gain at position k (NDCG@k).