Single-pass online learning: performance, voting schemes and online feature selection Export

@inproceedings{citeulike:2243777, address = {New York, NY, USA}, author = {Carvalho, Vitor R. and Cohen, William W.}, booktitle = {KDD '06: Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining}, citeulike-article-id = {2243777}, citeulike-linkout-0 = {http://portal.acm.org/citation.cfm?id=1150402.1150466}, citeulike-linkout-1 = {http://dx.doi.org/10.1145/1150402.1150466}, citeulike-linkout-2 = {http://www.cs.cmu.edu/~wcohen/postscript/kdd2006.pdf}, comment = {=== Pseudocodes and parameter settings === see inside paper === Online Feature Selection === Named in paper as: External Feature Selection (EFS) Ranks the feature importance according to the difference (in absolute value) between its positive and negative Balanced Winnow weights. More specifically, at each time t the importance score I of the feature j is given by I\_j = |u\_j - v\_j |, where u\_ and v\_j are the positive and the negative model weights for feature j. EFS uses not only the extreme top T features, but also a small number from the extreme bottom B. For instance, in order to select 100 features from a dataset, EFS would select 90\% of these 100 features from the extreme top T and 10\% from the extreme bottom B. It has been measured that EFS has a performance comparable to IG and CHI. }, doi = {10.1145/1150402.1150466}, isbn = {1595933395}, keywords = {averaging, balanced-winnow, feature-selection, online-feature-selection, wide-margin, winnow}, pages = {548--553}, posted-at = {2008-12-03 20:34:42}, priority = {5}, publisher = {ACM}, title = {Single-pass online learning: performance, voting schemes and online feature selection}, url = {http://dx.doi.org/10.1145/1150402.1150466}, year = {2006} }

TY - CONF ID - citeulike:2243777 L3 - citeulike-article-id:2243777 SN - 1595933395 AD - New York, NY, USA EP - 553 TI - Single-pass online learning: performance, voting schemes and online feature selection PB - ACM T2 - KDD '06: Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining SP - 548 KW - averaging KW - balanced-winnow KW - feature-selection KW - online-feature-selection KW - wide-margin KW - winnow N1 - === Pseudocodes and parameter settings === see inside paper === Online Feature Selection === Named in paper as: External Feature Selection (EFS) Ranks the feature importance according to the difference (in absolute value) between its positive and negative Balanced Winnow weights. More specifically, at each time t the importance score I of the feature j is given by I_j = |u_j - v_j |, where u_ and v_j are the positive and the negative model weights for feature j. EFS uses not only the extreme top T features, but also a small number from the extreme bottom B. For instance, in order to select 100 features from a dataset, EFS would select 90% of these 100 features from the extreme top T and 10% from the extreme bottom B. It has been measured that EFS has a performance comparable to IG and CHI. AU - Carvalho, Vitor AU - Cohen, William PY - 2006/// UR - http://dx.doi.org/10.1145/1150402.1150466 ER -