Learning to Rank for Information Retrieval Using the Clonal Selection Algorithm*

摘要

One central problem of Information Retrieval (IR) is ranking. This paper focuses on learning to rank for IR, which exploits machine learning to handle the ranking task. One state of the art approach for learning to rank is the listwise one, which uses document lists as "instance" in learning, and minimizes a loss function defined on the predicted permutation and the ground-truth permutation. In this paper we propose a novel listwise method, ListClonal, to address learning to rank for IR. ListClonal employs the clonal selection algorithm to learn an effective ranking function by combining various types of evidences in IR. A weighted rank measure of correlation called Shieh/b is introduced as the listwise loss function to precisely measure the rank performance for learning. Experimental results on the LETOR benchmark datasets show that ListClonal outperforms the baseline methods of BM25, LMIR, Ranking SVM, and AdaRank in terms of Kendall tau, Spearman's and Shieh/b rank correlation.