Abstract:

Designing effective ranking functions is a core problem for information retrieval since the ranking functions directly impacted the relevance of the search results. Learning ranking functions from preference data in particular have recently attracted much interest. The ranking algorithms were often evaluated using information retrieval measures. The main difficulty in direct optimization of these measures was that they depended on the ranks of documents. So it was important to optimize the ranking positions of relevant documents in the result list. Specifically, the roles of preference were investigated between the relevant documents and irrelevant documents in the learning process. To remedy this, a new input sample named one-group sample was constructed by a relevant document and a group of irrelevant documents according to a given query. The new sample could effectively distinguish the relevance of documents.  With the new samples a new position based loss function was also developed to improve the performance of learned ranking functions. Experimental studies were conducted using the Letor30 data set which improved ranking accuracies by 2％ and demonstrated the effectiveness of the proposed method.

Key words: learning to rank, information retrieval, ranking positions