12 Jul
Optimizing Search Engines using Clickthrough Data
PubDate(2002), PubPlace(SIGKDD) Author(Joachims)
keyword(SVM,Clickthrough,Pairwise preference,Learning to rank,Metasearch)
Content
Background
- LTR based on expert-judged relevance
Contribution
- Generation pairwise pref. using clickthrough
- Training this binary ordering using SVM
- By minimizing rank correlation to optimal ranking via Kendall’s
- By minimizing rank correlation to optimal ranking via Kendall’s
- Meta-search (Strive) engine to compare the learned ranking function with existing methods
- To perform unbiased comparison of different rankings with clickthrough data, rank combination method that equally presents the links from each system.
Experiment
- As more data was used for training, the algorithm showed lower rate of error.
Future Work
Comment
The author used the search engine himself to verify his result. I may need to build metasearch engine myself, which can be useful for a variety of tasks.
Reference
- http://www.joachims.org/publications/joachims_02c.pdf
- http://ciir.cs.umass.edu/wiki/bin/view/Main/IRseminar08-0215