Parallel Classifcation Algorithms And Their Applications To Text Classifcation

Machine learning has been applied to many felds like speech recognition, im-age recognition and text classifcation. Traditional pattern classifcation algorithms arealmost serial algorithms. Since the data grow larger and larger and the single coreCPU meets the limit, the serial algorithms become inefcient tand even inefective tohandle large-scale data. Nowadays the research of parallel classifcation algorithms isimportant.Ourresearchworkofparallelclassifcationalgorithmsbasesonmin-maxmodularneural network. Min-max modular neural network is a generalized modular learningframework whose feature is the parallel and modular struture. We study two classif-cation algorithms that based on min-max modular neural network: min-max modularneural network with gaussian zero-crossing functions (M3-GZC) and min-max modu-lar LIBLINEAR (M3-LIBLINEAR).M3-GZC is an instance-based learning algorithm which has the ability to output“unknown”. The ability to output “unknown” is important in some tasks which havehigh cost. In our previous work, we proposed a structure pruning algorithm to speedup M3-GZC by removing the redundant modules. Now we propose an approximatepruning algorithm to remove more redundant modules and make M3-GZC faster. Wealso put M3-GZC in MapReduce and design experiments to confrm the speedup.Tohandletextclassifcationproblem,weproposeM3-LIBLINEAR.M3-LIBLINEARisamin-maxmodularneuralnetworkwithLIBLINEARasitsbaseclassifer. Wecom-pare M3-LIBLINEAR with LIBLINEAR-CDBLOCK on NTCIR-5Japanese patentclassifcation problem, which is a large-scale multi-label imbalanced problem. Theresult shows that M3-LIBLINEAR has faster training speed and better generalization performance.