Taxonomies of the Web typically have hundreds of thousands of categories and skewed category distribution over documents. It is not clear whether existing text classification technologies can perform well on and scale up to such large-scale applications. To understand this, we conducted the evaluation of several representative methods (Support Vector Machines, <i>k</i>-Nearest Neighbor and Naive Bayes) with Yahoo! taxonomies. In particular, we evaluated the effectiveness/efficiency tradeoff in classifiers with hierarchical setting compared to conventional (flat) setting, and tested popular threshold tuning strategies for their scalability and accuracy in large-scale classification problems.

@inproceedings{liu2005experimental, author = {LIU, Tie-Yan and YANG, Yiming and WAN, Hao and ZHOU, Qian and GAO, Bin and ZENG, Hua-Jun and CHEN, Zheng and MA, Wei-Ying}, title = {An experimental study on large-scale web categorization}, booktitle = {Special interest tracks and posters of the 14th international conference on World Wide Web}, series = {WWW '05}, publisher = {ACM}, address = {New York, NY, USA}, year = {2005}, pages = {1106--1107}, url = {http://doi.acm.org/10.1145/1062745.1062891}, doi = {10.1145/1062745.1062891}, isbn = {1-59593-051-5}, keywords = {categorization, bachelor:2011:bachmann, web}, abstract = {Taxonomies of the Web typically have hundreds of thousands of categories and skewed category distribution over documents. It is not clear whether existing text classification technologies can perform well on and scale up to such large-scale applications. To understand this, we conducted the evaluation of several representative methods (Support Vector Machines, <i>k</i>-Nearest Neighbor and Naive Bayes) with Yahoo! taxonomies. In particular, we evaluated the effectiveness/efficiency tradeoff in classifiers with hierarchical setting compared to conventional (flat) setting, and tested popular threshold tuning strategies for their scalability and accuracy in large-scale classification problems.} }

%0 = inproceedings %A = LIU, Tie-Yan and YANG, Yiming and WAN, Hao and ZHOU, Qian and GAO, Bin and ZENG, Hua-Jun and CHEN, Zheng and MA, Wei-Ying %B = Special interest tracks and posters of the 14th international conference on World Wide Web %C = New York, NY, USA %D = 2005 %I = ACM %T = An experimental study on large-scale web categorization %U = http://doi.acm.org/10.1145/1062745.1062891

@inproceedings{kubica-stochastic, author = {Kubica, Jeremy and Moore, Andrew and Schneider, Jeff and Yang, Yiming}, title = {Stochastic Link and Group Detection}, booktitle = {Proceedings of the Eighteenth National Conference on Artificial Intelligence}, publisher = {AAAI Press/MIT Press}, year = {2002}, pages = {798--804}, keywords = {detection, gda, community, network} }

%0 = inproceedings %A = Kubica, Jeremy and Moore, Andrew and Schneider, Jeff and Yang, Yiming %B = Proceedings of the Eighteenth National Conference on Artificial Intelligence %D = 2002 %I = AAAI Press/MIT Press %T = Stochastic Link and Group Detection