Active Learning for Learning to Rank

• Rodrigo M. Silva UFMG
• Marcos A. Gonçalves UFMG
• Adriano Veloso UFMG

Abstract

This paper summarizes our master’s dissertation which proposes a novel method for actively sampling query-document instances from a document collection for labeling. Using our technique, it is possible to select and label a small and yet highly effective set that can be used to train Learning to Rank (L2R) algorithms. We conducted extensive experimentation of the method using benchmarking datasets to show that it obtains state-of-the-art results when compared to active and supervised baselines.

References

De Sousa, D. X., Rosa, T. C., Martins, W. S., Silva, R., and Gonçalves, M. A. (2012). Improving on-demand learning to rank through parallelism. In WISE’12, pages 526–537.

Donmez, P. and Carbonell, J. G. (2008). Optimizing estimated loss reduction for active sampling in rank learning. In ICML ’08, pages 248–255.

Ferreira, A. A., Silva, R., Gonçalves, M. A., Veloso, A., and Laender, A. H. (2012). Active associative sampling for author name disambiguation. In JCDL ’12, pages 175–184.

Silva, R., Gonçalves, M. A., and Veloso, A. (2011). Rule-based active sampling for learning to rank. In ECML PKDD ’11, pages 240–255.

Sumbana, M., Gonçalves, M. A., Silva, R., Almeida, J., and Veloso, A. (2012). Automatic vandalism detection in wikipedia with active associative classification. In TPDL’12, pages 138–143.

Veloso, A. A., Almeida, H. M., Gonçalves, M. A., and Meira, Jr., W. (2008). Learning to rank at query-time using association rules. In SIGIR ’08, pages 267–274.