SoftRank: optimizing non-smooth rank metrics Export

@inproceedings{citeulike:5835532, abstract = {We address the problem of learning large complex ranking functions. Most IR applications use evaluation metrics that depend only upon the ranks of documents. However, most ranking functions generate document scores, which are sorted to produce a ranking. Hence IR metrics are innately non-smooth with respect to the scores, due to the sort. Unfortunately, many machine learning algorithms require the gradient of a training objective in order to perform the optimization of the model parameters,and because IR metrics are non-smooth,we need to find a smooth proxy objective that can be used for training. We present a new family of training objectives that are derived from the rank distributions of documents, induced by smoothed scores. We call this approach SoftRank. We focus on a smoothed approximation to Normalized Discounted Cumulative Gain (NDCG), called SoftNDCG and we compare it with three other training objectives in the recent literature. We present two main results. First, SoftRank yields a very good way of optimizing NDCG. Second, we show that it is possible to achieve state of the art test set NDCG results by optimizing a soft NDCG objective on the training set with a different discount function}, address = {New York, NY, USA}, author = {Taylor, Michael and Guiver, John and Robertson, Stephen and Minka, Tom}, booktitle = {WSDM '08: Proceedings of the international conference on Web search and web data mining}, citeulike-article-id = {5835532}, citeulike-linkout-0 = {http://portal.acm.org/citation.cfm?id=1341544}, citeulike-linkout-1 = {http://dx.doi.org/10.1145/1341531.1341544}, doi = {10.1145/1341531.1341544}, isbn = {978-1-59593-927-9}, keywords = {learning, list, ltr, rank}, location = {Palo Alto, California, USA}, pages = {77--86}, posted-at = {2009-09-24 12:17:54}, priority = {2}, publisher = {ACM}, title = {SoftRank: optimizing non-smooth rank metrics}, url = {http://dx.doi.org/10.1145/1341531.1341544}, year = {2008} }

TY - CONF ID - citeulike:5835532 L3 - citeulike-article-id:5835532 N2 - We address the problem of learning large complex ranking functions. Most IR applications use evaluation metrics that depend only upon the ranks of documents. However, most ranking functions generate document scores, which are sorted to produce a ranking. Hence IR metrics are innately non-smooth with respect to the scores, due to the sort. Unfortunately, many machine learning algorithms require the gradient of a training objective in order to perform the optimization of the model parameters,and because IR metrics are non-smooth,we need to find a smooth proxy objective that can be used for training. We present a new family of training objectives that are derived from the rank distributions of documents, induced by smoothed scores. We call this approach SoftRank. We focus on a smoothed approximation to Normalized Discounted Cumulative Gain (NDCG), called SoftNDCG and we compare it with three other training objectives in the recent literature. We present two main results. First, SoftRank yields a very good way of optimizing NDCG. Second, we show that it is possible to achieve state of the art test set NDCG results by optimizing a soft NDCG objective on the training set with a different discount function CY - Palo Alto, California, USA SN - 978-1-59593-927-9 AD - New York, NY, USA EP - 86 TI - SoftRank: optimizing non-smooth rank metrics PB - ACM T2 - WSDM '08: Proceedings of the international conference on Web search and web data mining SP - 77 KW - learning KW - list KW - ltr KW - rank AU - Taylor, Michael AU - Guiver, John AU - Robertson, Stephen AU - Minka, Tom PY - 2008/// UR - http://dx.doi.org/10.1145/1341531.1341544 ER -