Expected Loss Optimization for Document Ranking by Active Learning

Learning to rank is the emerging research field in many data mining applications and information retrieval techniques (e.g. Search engines). The major issue in ranking algorithm is that the quality or ranking is affected by labeled examples, since it is very expensive and also time consuming to collect labeled samples. This problem brings a great need for active learning algorithm; however, in literature learning to rank uses supervised learning algorithm where ranking is based on labeled data only. A general active learning framework Balanced two stage Expected Loss Optimization is proposed to select the most informative document based on user’s query. The algorithm is based on two levels, Query level and Document level and grade distribution is done based on query and document pairs. Experiment on web search dataset has demonstrated with the proposed algorithm.