Ranking under Tight Budgets
Christian Pölitz, Ralf Schenkel · 2012
This paper introduces a budget-aware learning to rank approach that limits the cost for evaluating a ranking model, with a focus on very tight budgets that do not allow to fully evaluate at least for one time all documents for each term. In contrast to existing work on budget-aware learning to rank, our model allows to only partially evaluate parts of the ranking model for the most promising documents. In contrast to existing work on top-k retrieval, we generate an execution plan before the actual query processing starts, eliminating the need for expensive in-memory accumulator management. We consider a unified cost model that integrates loading and processing cost. An extensive evaluation with a standard benchmark collection shows that our method outperforms other budget-aware methods under tight budgets in terms of result quality.