21.13 Remove English Stop Words
docs <- tm_map(docs, removeWords, stopwords("english"))inspect(docs[16])## <<SimpleCorpus>>
## Metadata: corpus specific: 1, document level (indexed): 0
## Content: documents: 1
##
## hwrf12.txt
## hybrid weighted random forests \nclassifying highdimensional data\nbaoxun xu joshua zhexue huang graham williams \nyunming ye\n\n\ndepartment computer science harbin institute technology shenzhen graduate\nschool shenzhen china\n\nshenzhen institutes advanced technology chinese academy sciences shenzhen\n china\nemail amusing gmailcom\nrandom forests popular classification method based ensemble \nsingle type decision trees subspaces data literature \n many different types decision tree algorithms including c cart \nchaid type decision tree algorithm may capture different information\n structure paper proposes hybrid weighted random forest algorithm\nsimultaneously using feature weighting method hybrid forest method \nclassify high dimensional data hybrid weighted random forest algorithm\ncan effectively reduce subspace size improve classification performance\nwithout increasing error bound conduct series experiments eight\nhigh dimensional datasets compare method traditional random forest\nmethods classification methods results show method\nconsistently outperforms traditional methods\nkeywords random forests hybrid weighted random forest classification decision tree\n\n\n\nintroduction\n\nrandom forests popular classification\nmethod builds ensemble single type\n decision trees different random subspaces \ndata decision trees often either built using\nc cart one type within\n single random forest recent years random\nforests attracted increasing attention due \n competitive performance compared \nclassification methods especially highdimensional\ndata algorithmic intuitiveness simplicity \n important capability ensemble using\nbagging stochastic discrimination \nseveral methods proposed grow random\nforests subspaces data \n methods popular forest construction\nprocedure proposed breiman first use\nbagging generate training data subsets building\nindividual trees\n subspace features \nrandomly selected node grow branches \n decision tree trees combined \nensemble forest ensemble learner \nperformance random forest highly dependent\n two factors performance tree \ndiversity trees forests breiman\nformulated overall performance set trees \n average strength proved generalization\n\nerror random forest bounded ratio \naverage correlation trees divided square\n average strength trees\n high dimensional data text data\n usually large portion features \nuninformative classes forest building\nprocess informative features large\nchance missed randomly select small\nsubspace breiman suggested selecting log m \nfeatures subspace m number \nindependent features data high dimensional\ndata result weak trees created \nsubspaces average strength trees reduced\n error bound random forest enlarged\ntherefore large proportion weak\ntrees generated random forest forest \nlarge likelihood make wrong decision mainly\nresults weak trees classification power\n address problem aim optimize decision\ntrees random forest two strategies one\nstraightforward strategy enhance classification\nperformance individual trees feature weighting\nmethod subspace sampling \nmethod feature weights computed respect\n correlations features class feature\n regarded probabilities feature \n selected subspaces method obviously\nincreases classification performance individual\n\n computer journal vol \n\n \n\n\n\n\n\nbaoxun xu joshua zhexue huang graham williams yunming ye\n\ntrees subspaces will biased contain\n informative features however chance \ncorrelated trees also increased since features \nlarge weights likely repeatedly selected\n second strategy straightforward use\nseveral different types decision trees training\ndata subset increase diversity trees\n select optimal tree individual\ntree classifier random forest model work\npresented extends algorithm developed \nspecifically build three different types tree\nclassifiers c cart chaid \ntraining data subset evaluate performance\n three classifiers select best tree \n way build hybrid random forest may\ninclude different types decision trees ensemble\n added diversity decision trees can effectively\nimprove accuracy tree forest \nhence classification performance ensemble\nhowever use method build best\nrandom forest model classifying high dimensional\ndata can sure subspace size best\n paper propose hybrid weighted random\nforest algorithm simultaneously using new feature\nweighting method together hybrid random\nforest method classify high dimensional data \n new random forest algorithm calculate feature\nweights use weighted sampling randomly select\nfeatures subspaces node building different\ntypes trees classifiers c cart chaid \n training data subset select best tree \n individual tree final ensemble model\nexperiments performed high dimensional\ntext datasets dimensions ranging \n compared performance eight random\nforest methods wellknown classification methods\nc random forest cart random forest chaid\nrandom forest hybrid random forest c weighted\nrandom forest cart weighted random forest chaid\nweighted random forest hybrid weighted random\nforest support vector machines naive bayes \n knearest neighbors \n experimental\nresults show hybrid weighted random forest\nachieves improved classification performance \nten competitive methods\n remainder paper organized follows\n section introduce framework building\n hybrid weighted random forest describe new\nrandom forest algorithm section summarizes four\nmeasures evaluate random forest models present\nexperimental results high dimensional text datasets\n section section contains conclusions\n\ntable contingency table input feature class\nfeature y\ny y \ny yj \ny yq total\n \n\n\nj\n\nq\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n ai\n\n\nij\n\niq\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n ap\np\n\npj\n\npq\np\ntotal\n\n\nj\n\nq\n\n\ngeneral framework building hybrid random forests\n integrating two methods propose novel\nhybrid weighted random forest algorithm\n\n\nlet y class target feature q distinct\nclass labels yj j q purposes \n discussion consider single categorical feature\n dataset d p distinct category values \ndenote distinct values ai p\nnumeric features can discretized p intervals \n supervised discretization method\nassume d val objects size subset \nd satisfying condition ai y yj \ndenoted ij considering combinations \ncategorical values labels y can\nobtain contingency table y shown\n table far right column contains marginal\ntotals feature \n\nhybrid\nforests\n\nweighted\n\nrandom\n\n section first introduce feature weighting\nmethod subspace sampling present \n\nq\n\n\n \n\nij\n\n p\n\n\n\nj\n\n bottom row marginal totals class\nfeature y \nj \n\np\n\n\nij\n\n j q\n\n\n\n\n\n grand total total number samples \n bottom right corner\n\n\nq \np\n\n\nij\n\n\n\nj \n\ngiven training dataset d feature first\ncompute contingency table feature weights \n computed using two methods discussed\n following subsection\n\n\n\n\nnotation\n\nfeature weighting method\n\n subsection give details feature\nweighting method subspace sampling random\nforests consider mdimensional feature space\n present compute \n\n computer journal vol \n\n \n\n\n\nhybrid weighted random forests classifying highdimensional data\nweights w w wm every feature space\n weights used improved algorithm\n grow decision tree random forest\n feature weight computation\n weight feature represents correlation\n values feature values \nclass feature y larger weight will indicate \nclass labels objects training dataset \ncorrelated values feature indicating \n informative class objects thus \n suggested stronger power predicting\n classes new objects\n following propose use chisquare\nstatistic compute feature weights \nmethod can quantify correspondence two\ncategorical variables\ngiven contingency table input feature \n class feature y dataset d chisquare statistic\n two features computed \ncorra y \n\nq\np \n\nij tij \ntij\n j\n\n\n\n ij observed frequency \ncontingency table tij expected frequency\ncomputed \n x j\ntij \n\n\n\n\n larger measure corra y \ninformative feature predicting class y \n normalized feature weight\n practice feature weights normalized feature\nsubspace sampling use corra y measure \ninformativeness features consider \n feature weights however treat weights \nprobabilities features normalize measures \nensure sum normalized feature weights \nequal let corrai y m set\n m feature measures compute normalized\nweights \n\ncorrai y \nwi n \n\n corrai y \n use square root smooth values \n measures wi can considered probability\n feature ai randomly sampled subspace \n informative feature larger weight \n higher probability feature selected\n\ndiversity commonly obtained using bagging \nrandom subspace sampling introduce \nelement diversity using different types trees\nconsidering analogy forestry different data subsets bagging represent soil structures different decision tree algorithms represent different tree species approach two key aspects\none use three types decision tree algorithms \ngenerate three different tree classifiers training data subset evaluate accuracy\n tree measure tree importance \npaper use outofbag accuracy assess importance tree\nfollowing breiman use bagging generate\n series training data subsets build\ntrees tree data subset used grow\n tree called inofbag iob data \nremaining data subset called outofbag oob\ndata since oob data used building trees\n can use data objectively evaluate trees\naccuracy importance oob accuracy gives \nunbiased estimate true accuracy model\ngiven n instances training dataset d tree\nclassifier hk iobk built kth training data\nsubset iobk define oob accuracy tree\nhk iobk di d \nn\noobacck \n\nframework building hybrid random\nforest\n\n ensemble learner performance random\nforest highly dependent two factors diversity\namong trees accuracy tree \n\n\n\nihk di yi di oobk \nn\n idi oobk \n\n\n\n indicator function larger \noobacck better classification quality tree\n use outofbag data subset oobi calculate\n outofbag accuracies three types trees\nc cart chaid evaluation values e \ne e respectively\nfig illustrates procedure building hybrid\nrandom forest model firstly series iob oob\ndatasets generated entire training dataset\n bagging three types tree classifiers c\ncart chaid built using iob dataset\n corresponding oob dataset used calculate \noob accuracies three tree classifiers finally\n select tree highest oob accuracy \n final tree classifier included hybrid\nrandom forest\nbuilding hybrid random forest model \nway will increase diversity among trees\n classification performance individual tree\nclassifier also maximized\n\n\n\n\n\n\ndecision tree algorithms\n\n core approach diversity decision\ntree algorithms random forest different decision\ntree algorithms grow structurally different trees \n training data selecting good decision tree\nalgorithm grow trees random forest critical\n\n computer journal vol \n\n \n\n\n\n\n\nbaoxun xu joshua zhexue huang graham williams yunming ye\n difference lies way split node \n split functions binary branches multibranches work use different decision\ntree algorithms build hybrid random forest\n\n\n\nfigure hybrid random forests framework\n\n performance random forest studies\n considered different decision tree algorithms\naffect random forest paper\n common decision tree algorithms follows\nclassification trees c supervised\nlearning classification algorithm used construct\ndecision trees given set preclassified objects \ndescribed vector attribute values construct\n mapping attribute values classes c uses\n divideandconquer approach grow decision trees\nbeginning entire dataset tree constructed\n considering predictor variable dividing \ndataset best predictor chosen node\nusing impurity diversity measure goal \n produce subsets data homogeneous\n respect target variable c selects test\n maximizes information gain ratio igr \nclassification regression tree cart \n recursive partitioning method can used \n regression classification main difference\n c cart test selection \nevaluation process\nchisquared automatic interaction detector\nchaid method based chisquare test \nassociation chaid decision tree constructed\n repeatedly splitting subsets space two\n nodes determine best split \nnode allowable pair categories predictor\nvariables merged statistically\nsignificant difference within pair respect \ntarget variable \n decision tree algorithms can see \n\nhybrid weighted random forest algorithm\n\n subsection present hybrid weighted\nrandom forest algorithm simultaneously using \nfeature weights hybrid method classify high\ndimensional data benefits algorithm \ntwo aspects firstly compared hybrid forest\nmethod can use small subspace size \ncreate accurate random forest models\nsecondly\ncompared building random forest using feature\nweighting can use several different types \ndecision trees training data subset increase\n diversities trees added diversity \ndecision trees can effectively improve classification\nperformance ensemble model detailed steps\n introduced algorithm \ninput parameters algorithm include training\ndataset d set features class feature y \n number trees random forest k \nsize subspaces m output random forest\nmodel m lines form loop building k\ndecision trees loop line samples training\ndata d sampling replacement generate \ninofbag data subset iobi building decision tree\nline build three types tree classifiers c\ncart chaid procedure line calls\n function createt reej build tree classifier\nline calculates outofbag accuracy tree\nclassifier procedure line selects tree\nclassifier maximum outofbag accuracy k\ndecision tree trees thus generated form hybrid\nweighted random forest model m \ngenerically function createt reej first creates \nnew node tests stopping criteria decide\nwhether return upper node split \nnode choose split node feature\nweighting method used randomly select m features\n subspace node splitting features\n used candidates generate best split \npartition node subset partition\ncreatet reej called create new node \n current node leaf node created returns \n parent node recursive process continues \n full tree generated\n\n computer journal vol \n\n \n\n\n\nhybrid weighted random forests classifying highdimensional data\nalgorithm new random forest algorithm\n input\n d training dataset\n features space \n y class features space y y yq \n k number trees\n m size subspaces\n output random forest m \n method\n k \n\ndraw bootstrap sample inofbag data subset\niobi outofbag data subset oobi \ntraining dataset d\n\n j \n\nhij iobi createt reej \nuse outofbag data subset oobi calculate\n\n outofbag accuracy oobacci j tree\nclassifier hij iobi equation\n\nend \n\nselect hi iobi highest outofbag\naccuracy oobacci optimal tree \n end \n combine\n\nk\ntree\nclassifiers\nh iob h iob hk iobk random\nforest m \n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nfunction createtree\ncreate new node n \n stopping criteria met \nreturn n leaf node\nelse\n j m \ncompute\n\ninformativeness\nmeasure\ncorraj y equation \nend \ncompute feature weights w w wm \nequation \nuse feature weighting method randomly\nselect m features\nuse m features candidates generate\n best split node partitioned\ncall createtree split\nend \nreturn n \nevaluation measures\n\n paper use five measures ie strength\ncorrelation error bound c s test accuracy f\nmetric evaluate random forest models strength\nmeasures collective performance individual trees\n random forest correlation measures \ndiversity trees ratio correlation\n square strength c s indicates \ngeneralization error bound random forest model\n three measures introduced \naccuracy measures performance random forest\nmodel unseen test data f metric \n\n\n\ncommonly used measure classification performance\n\n\nstrength correlation measures\n\n follow breimans method described \ncalculate strength correlation ratio c s \nfollowing breimans notation denote strength \ns correlation let hk iobk kth\ntree classifier grown kth training data iobk\nsampled d replacement\nassume \nrandom forest model contains k trees outofbag\nproportion votes di d class j \nk\nihk di j di \n iobk \nqdi j kk\n\n iobk \nk idi \n number trees random forest\n trained without di classify di class\nj divided number training datasets \ncontaining di \n strength s computed \n\nqdi yi maxjyi qdi j\nn \nn\n\ns\n\n\n\n n number objects d yi indicates\n true class di \n correlation computed \nn\n\n\n\n qdi yi maxjyi qdi j s\nn\n\n \n\n\nk\n\nk\nk pk pk \nk pk p\n\n\nn\npk \n\n\n\nihk di yi di \n iobk \nn\n iobk \n idi \n\n\n\n\nn\npk \n\n\n\nihk di jdi y di \n iobk \nn\nid\n\n \niob\n\n\nk\n\n\n\n\n\njdi y argmaxjyi qd j\n\n\n\n class obtains maximal number votes\namong classes true class\n\n\ngeneral error bound measure c s\n\ngiven strength correlation outofbag\nestimate c s measure can computed\n important theoretical result breimans method\n upper bound generalization error \nrandom forest ensemble derived \np e s s\n\n\n\n mean value correlations \npairs individual classifiers s strength \n set individual classifiers estimated \n\n computer journal vol \n\n \n\n\n\n\n\nbaoxun xu joshua zhexue huang graham williams yunming ye\n\naverage accuracy individual classifiers d \noutofbag evaluation inequality shows \ngeneralization error random forest affected \n strength individual classifiers mutual\ncorrelations therefore breiman defined c s ratio\n measure random forest \nc s s\n\n\n\n smaller ratio better performance \n random forest c s gives guidance \nreducing generalization error random forests\n\n\ntest accuracy\n\n test accuracy measures classification performance random forest test data set let\ndt test data yt class labels given\ndi dt number votes di class j \nn di j \n\nk\n\n\nihk di j\n\n\n\ntable \nsummary statistic highdimensional\ndatasets\nname\nfeatures\ninstances\nclasses minority\nfbis\n\n\n\n\nre\n\n\n\n\nre\n\n\n\n\ntr\n\n\n\n\nwap\n\n\n\n\ntr\n\n\n\n\nlas\n\n\n\n\nlas\n\n\n\n\n\n emphasizes performance classifier rare\ncategories define follows\n\n \n\nt pi\nt pi\n \nt pi f pi \nt pi f ni \n\n\n\nf category macroaveraged f\n computed \n\nk\n\n test accuracy calculated \nf \n\n di yi maxjyi n di j \nn \n\n \n m acrof \n \n\nq\n\n\nq\n\nf \n\n\n\nn\n\nacc \n\n n number objects dt yi indicates\n true class di \n\n\nf metric\n\n evaluate performance classification methods\n dealing unbalanced class distribution use\n f metric introduced yang liu \nmeasure equal harmonic mean recall \n precision overall f score entire\nclassification problem can computed microaverage macroaverage\nmicroaveraged f computed globally \nclasses emphasizes performance classifier\n common classes define follows\nq\n\nq\nt pi\n t pi\n q \n q\n\n t pi f pi \n t pi f ni \n q number classes t pi true positives\n number objects correctly predicted class \nf pi false positives number objects \npredicted belong class microaveraged f computed \nm icrof \n\n\n\n\n\n\nmacroaveraged f first computed locally \n class average classes taken\n\n larger microf macrof values \nhigher classification performance classifier\n\n\nexperiments\n\n section present two experiments \ndemonstrate effectiveness new random\nforest algorithm classifying high dimensional data\nhigh dimensional datasets various sizes \ncharacteristics used experiments \nfirst experiment designed show proposed\nmethod can reduce generalization error bound\nc s improve test accuracy size \n selected subspace large second\nexperiment used demonstrate classification\nperformance proposed method comparison \n classification methods ie svm nb knn\n\n\ndatasets\n\n experiments used eight realworld high\ndimensional datasets datasets selected\ndue diversities number features \nnumber instances number classes \ndimensionalities vary instances\nvary minority class rate varies\n dataset randomly\nselect instances training dataset \n remaining data test dataset detailed\ninformation eight datasets listed table \n fbis re re tr wap tr las\n las datasets classical text classification\nbenchmark datasets carefully selected \n\n computer journal vol \n\n \n\n\n\nhybrid weighted random forests classifying highdimensional data\npreprocessed han karypis dataset fbis\n compiled foreign broadcast information\nservice trec datasets re re \nselected reuters text categorization test\ncollection distribution datasets tr \ntr derived trec trec \n trec dataset wap webace\nproject wap datasets las las \nselected los angeles times trec \n classes datasets generated \nrelevance judgment provided collections\n\n\nperformance comparisons random forest methods\n\n purpose experiment evaluate\n effect hybrid weighted random forest\nmethod h w rf strength correlation c s \n test accuracy\n eight high dimensional\ndatasets analyzed results compared\n seven random forest methods ie c\nrandom forest c rf cart random forest\ncart rf chaid random forest chaid rf\nhybrid random forest h rf c weighted random\nforest c w rf cart weighted random forest\ncart w rf chaid weighted random forest\nchaid w rf dataset ran \nrandom forest algorithm different sizes \nfeature subspaces since number features \ndatasets large started subspace\n features increased subspace \nfeatures time given subspace size built\n trees random forest model order \nobtain stable result built random forest models\n subspace size dataset algorithm\n computed average values four measures\n strength correlation c s test accuracy \nfinal results comparison performance \neight random forest algorithms four measures\n datasets shown figs \n\nfig plots strength eight methods \ndifferent subspace sizes datasets\n subspace higher strength \nbetter result curves can see \n new algorithm h w rf consistently performs\nbetter seven random forest algorithms\n advantages obvious small subspaces\n new algorithm quickly achieved higher strength\n subspace size increases\n seven \nrandom forest algorithms require larger subspaces \nachieve higher strength results indicate \n hybrid weighted random forest algorithm enables\nrandom forest models achieve higher strength\n small subspace sizes compared seven \nrandom forest algorithms\nfig plots curves correlations \neight random forest methods datasets \n\n\n\nsmall subspace sizes h rf c rf cart rf\n chaid rf produce higher correlations \n trees datasets correlation decreases\n subspace size increases random forest\nmodels lower correlation trees\n better final model\n new\nrandom forest algorithm h w rf low correlation\nlevel achieved small subspaces \n datasets also note subspace size\nincreased correlation level increased well \nunderstandable subspace size increases\n informative features likely \nselected repeatedly subspaces increasing \nsimilarity decision trees therefore feature\nweighting method subspace selection works well \nsmall subspaces least point view \ncorrelation measure\nfig shows error bound indicator c s \neight methods datasets figures\n can observe subspace size increases c s\nconsistently reduces behaviour indicates \nsubspace size larger log m benefits eight\nalgorithms however new algorithm h w rf\nachieved lower level c s subspace size \nlog m seven algorithms\nfig plots curves showing accuracy \neight random forest models test datasets \n datasets can clearly see new random\nforest algorithm h w rf outperforms seven\n random forest algorithms eight data sets\n can seen new method stable\n classification performance methods \n figures observed highest test\naccuracy often obtained default subspace size\n log m implies practice large\nsize subspaces necessary grow highquality\ntrees random forests\n\n\nperformance comparisons\nclassification methods\n\n\n\n\n\n conducted experimental comparison\n three widely used text classification\nmethods support vector machines svm naive\nbayes nb knearest neighbor knn \nsupport vector machine used linear kernel \nregularization parameter often\nused text categorization naive bayes \nadopted multivariate bernoulli event model \n frequently used text classification knearest neighbor knn set number k \nneighbors experiments used wekas\nimplementation three text classification\nmethods used single subspace size \nfeatures eight datasets run random forest\nalgorithms h rf c rf cart rf \nchaid rf used subspace size features \n first datasets ie fbis re re tr wap \n\n computer journal vol \n\n \n\n\n\n\n\nbaoxun xu joshua zhexue huang graham williams yunming ye\nfbis\n\nre\n\n\n\n\n\n\n\n\n\n\n\nstrength\n\nstrength\n\n\n\n\n\n\n\nhwrf\n\n\n\ncwrf\ncartwrf\n\n\n\nhwrf\ncwrf\n\n\n\ncartwrf\n\nchaidwrf\n\nchaidwrf\n\n\n\nhrf\n\n\n\nhrf\n\ncrf\n\ncrf\n\n\n\ncartrf\n\n\n\ncartrf\n\nchaidrf\n\nchaidrf\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nnumber features\n\nnumber features\n\nre\n\n\n\n\n\ntr\n\n\n\n\n\n\n\n\n\n\n\n\nhwrf\ncwrf\n\n\n\ncartwrf\n\nstrength\n\nstrength\n\n\n\nhwrf\ncwrf\n\n\n\ncartwrf\n\nchaidwrf\n\n\n\nchaidwrf\n\nhrf\n\nhrf\n\n\n\ncrf\n\n\n\ncrf\n\ncartrf\n\ncartrf\n\n\n\nchaidrf\n\nchaidrf\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nnumber features\n\nnumber features\n\nwap\n\ntr\n\n\n\n\n\n\n\n\n\n\nstrength\n\nhwrf\n\n\n\ncwrf\ncartwrf\n\n\n\nstrength\n\n\n\n\n\n\nhwrf\ncwrf\n\n\n\ncartwrf\nchaidwrf\n\nchaidwrf\nhrf\n\n\n\nhrf\n\n\n\ncrf\n\ncrf\n\ncartrf\n\ncartrf\n\n\n\n\n\nchaidrf\n\nchaidrf\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nlas\n\n\n\n\n\n\n\n\n\n\n\nlas\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nhwrf\n\n\n\ncwrf\ncartwrf\nchaidwrf\n\n\n\nstrength\n\nstrength\n\n\n\nnumber features\n\nnumber features\n\n\n\n\n\nhwrf\n\n\n\ncwrf\ncartwrf\n\n\n\nchaidwrf\n\nhrf\ncrf\n\n\n\nhrf\n\n\n\ncrf\n\ncartrf\nchaidrf\n\ncartrf\n\n\n\nchaidrf\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nnumber features\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nnumber features\n\nfigure strength changes number features subspace high dimensional datasets\n\ntr run random forest algorithms used\n subspace size features last datasets\nlas las run random forest algorithms\n h w rf c w rf cart w rf \nchaid w rf used breimans subspace size \n\nlog m run random forest algorithms\n number features provided consistent result \nshown fig order obtain stable results \nbuilt random forest models random forest\nalgorithm dataset present average\n\n computer journal vol \n\n \n\n\n\nhybrid weighted random forests classifying highdimensional data\nfbis\n\n\n\nre\n\n\n\n\n\n\n\n\n\ncorrelation\n\ncorrelation\n\n\n\n\n\n\n\n\nhwrf\ncwrf\n\n\n\ncartwrf\nchaidwrf\n\n\n\nhwrf\n\n\n\ncwrf\ncartwrf\n\n\n\nchaidwrf\n\nhrf\ncrf\n\n\n\nhrf\n\n\n\ncrf\n\ncartrf\nchaidrf\n\ncartrf\n\n\n\nchaidrf\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nnumber features\n\n\n\n\n\n\n\n\n\n\n\n\n\nnumber features\n\nre\n\n\n\n\n\ntr\n\n\n\n\n\n\n\n\n\nhwrf\ncwrf\n\n\n\ncartwrf\nchaidwrf\n\n\n\ncorrelation\n\ncorrelation\n\n\n\n\n\nhwrf\n\n\n\ncwrf\ncartwrf\n\n\n\nchaidwrf\nhrf\n\nhrf\ncrf\n\n\n\ncrf\n\n\n\ncartrf\n\ncartrf\n\n\n\nchaidrf\n\nchaidrf\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nnumber features\n\nnumber features\n\nwap\n\ntr\n\n\n\n\n\n\n\n\ncorrelation\n\n\n\n\n\nhwrf\ncwrf\n\n\n\ncartwrf\nchaidwrf\n\n\n\ncorrelation\n\n\n\n\n\nhwrf\ncwrf\n\n\n\ncartwrf\nchaidwrf\n\n\n\nhrf\n\nhrf\n\ncrf\n\n\n\ncrf\n\ncartrf\n\ncartrf\n\n\n\nchaidrf\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nchaidrf\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nnumber features\n\nnumber features\n\nlas\n\nlas\n\n\n\n\n\n\n\n\n\n\n\n\n\nhwrf\n\n\n\ncwrf\ncartwrf\n\n\n\nchaidwrf\n\n\n\ncorrelation\n\ncorrelation\n\n\n\n\n\n\n\nhwrf\ncwrf\n\n\n\ncartwrf\nchaidwrf\n\n\n\nhrf\n\nhrf\ncrf\n\n\n\ncrf\n\n\n\ncartrf\n\ncartrf\nchaidrf\n\nchaidrf\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nnumber features\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nnumber features\n\nfigure correlation changes number features subspace high dimensional datasets\n\nresults noting range values less \n hybrid trees always accurate\n comparison results classification performance\n eleven methods shown table \n\nperformance estimated using test accuracy acc\n\nmicro f mic macro f mac boldface\ndenotes best results eleven classification\nmethods\n improvement often quite\nsmall always improvement demonstrated\n observe proposed method h w rf\n\n computer journal vol \n\n \n\n\n\n\n\nbaoxun xu joshua zhexue huang graham williams yunming ye\nfbis\n\n\n\nre\n\n\n\n\nlog m\n\n\n\n\n\n\n\n\n\n\n\ncwrf\n\n\n\n\n\n\n\nhwrf\n\nc s\n\nc s\n\n\n\n\n\ncartwrf\n\nhwrf\ncwrf\n\n\n\ncartwrf\n\nchaidwrf\n\n\n\nchaidwrf\n\n\n\nhrf\n\nhrf\n\nlog m\n\n\ncrf\n\n\n\ncrf\n\n\n\ncartrf\n\ncartrf\n\nchaidrf\n\n\n\nchaidrf\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nnumber features\n\n\n\n\n\n\n\n\n\n\n\n\n\nlog m\n\n\n\n\n\n\n\nc s\n\nhwrf\n\n\n\n\n\n\n\n\n\n\n\ntr\n\n\n\n\n\nc s\n\n\n\nnumber features\n\nre\n\n\n\n\n\ncwrf\n\n\n\n\n\nhwrf\ncwrf\n\n\n\ncartwrf\n\ncartwrf\n\nchaidwrf\n\n\n\nchaidwrf\n\n\n\nhrf\n\nhrf\n\ncrf\n\nlog m\n\n\n\n\n\ncrf\n\n\n\ncartrf\n\ncartrf\n\nchaidrf\n\n\n\nchaidrf\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nnumber features\n\nnumber features\n\nwap\n\ntr\n\n\n\n\n\n\n\nlog m\n\nlog m\n\n\n\n\n\n\n\n\ncwrf\n\nc s\n\n\n\nc s\n\nhwrf\n\n\n\n\n\n\n\n\n\nhwrf\n\n\n\ncwrf\ncartwrf\n\ncartwrf\n\n\nchaidwrf\n\n\n\nchaidwrf\nhrf\n\nhrf\ncrf\n\n\n\ncrf\n\n\n\ncartrf\n\ncartrf\nchaidrf\n\nchaidrf\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nnumber features\n\nnumber features\n\nlas\n\nlas\n\n\n\n\n\n\n\n\n\n\n\ncwrf\ncartwrf\n\n\n\n\n\nhrf\n\n\n\ncrf\ncartwrf\n\n\n\nchaidrf\n\n\n\n\n\n\n\n\n\nhwrf\n\n\n\ncwrf\ncartwrf\n\n\n\nchaidwrf\n\nlog m\n\n\n\nc s\n\nc s\n\n\n\n\nhwrf\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nchaidwrf\n\nlog m\n\n\n\n\nhrf\ncrf\n\n\n\ncartrf\nchaidrf\n\n\n\n\n\n\n\n\n\nnumber features\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nnumber features\n\nfigure c s changes number features subspace high dimensional datasets\n\noutperformed classification methods \ndatasets\n\n\n\nconclusions\n\n paper presented hybrid weighted random\nforest algorithm simultaneously using feature\nweighting method hybrid forest method classify\n computer journal vol \n\n \n\n\n\nhybrid weighted random forests classifying highdimensional data\nfbis\n\n\n\n\n\nre\n\n\n\n\n\n\n\n\n\n\nhwrf\ncwrf\n\n\n\ncartwrf\n\naccuracy\n\naccuracy\n\n\n\nchaidwrf\nhrf\n\n\n\n\nhwrf\n\n\n\ncwrf\ncartwrf\n\n\n\nchaidwrf\nhrf\n\n\n\ncrf\n\ncrf\n\nlog m\n\ncartrf\n\n\n\n\n\ncartrf\n\nlog m\n\n\n\n\n\nchaidrf\n\nchaidrf\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nnumber features\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nnumber features\n\nre\n\ntr\n\n\n\n\n\n\n\n\nlog m\n\n\n\n\n\n\n\n\nhwrf\ncwrf\n\n\n\ncartwrf\n\n\n\naccuracy\n\naccuracy\n\n\n\nchaidwrf\n\n\n\nhwrf\n\n\n\ncwrf\ncartwrf\n\n\n\nchaidwrf\nhrf\n\nhrf\n\n\n\nlog m\n\n\n\ncrf\n\ncrf\n\n\n\ncartrf\n\ncartrf\n\n\n\n\n\nchaidrf\n\nchaidrf\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nnumber features\n\nnumber features\n\nwap\n\n\n\n\n\ntr\n\n\n\n\n\n\n\n\n\n\n\n\n\nhwrf\ncwrf\n\n\n\ncartwrf\n\naccuracy\n\naccuracy\n\n\n\n\n\n\n\nhwrf\ncwrf\n\n\n\ncartwrf\nchaidwrf\n\nchaidwrf\n\n\n\nhrf\n\nlog m\n\n\n\n\nhrf\n\ncrf\n\n\n\ncartrf\n\n\n\ncrf\n\nlog m\n\n\n\n\ncartrf\nchaidrf\n\nchaidrf\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nnumber features\n\nnumber features\n\nlas\n\n\n\n\n\nlas\n\n\n\n\n\n\n\naccuracy\n\n\nhwrf\ncwrf\n\n\n\ncartwrf\nchaidwrf\n\n\n\naccuracy\n\n\n\n\n\n\n\nhwrf\ncwrf\ncartwrf\n\n\n\nchaidwrf\nhrf\n\nhrf\n\nlog m\n\n\n\ncrf\n\n\n\ncrf\n\nlog m\n\n\n\n\n\ncartrf\n\ncartrf\nchaidrf\n\nchaidrf\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nnumber features\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nnumber features\n\nfigure test accuracy changes number features subspace high dimensional datasets\n\nhigh dimensional data algorithm retains\n small subspace size breimans formula log m \n determining subspace size create accurate\nrandom forest models also effectively reduces\n upper bound generalization error \n\nimproves classification performance results \nexperiments various high dimensional datasets \nrandom forest generated new method superior\n classification methods can use default\nlog m subspace size generally guarantee\n\n computer journal vol \n\n \n\n\n\n\n\nbaoxun xu joshua zhexue huang graham williams yunming ye\n\ntable comparison results\ndatasets\ndataset\nfbis\nmeasures\nacc\nmic\nsvm\n \nknn\n\n\nnb\n \nh rf\n \nc rf\n \ncart rf\n \nchaid rf\n \nh w rf\n \nc w rf\n \ncart w rf\n \nchaid w rf\n \ndataset\nwap\nmeasures\nacc\nmic\nsvm\n\n\nknn\n \nnb\n \nh rf\n \nc rf\n \ncart rf\n \nchaid rf\n \nh w rf\n \nc w rf\n \ncart w rf\n\n\nchaid w rf\n \n\nbest accuracy micro f macro f results eleven methods \nre\nmic\n\n\n\n\n\n\n\n\n\n\n\ntr\nmac\nacc\nmic\n \n \n \n \n \n\n \n \n\n \n \n\n\n\n\n\n\n\nmac\n\n\n\n\n\n\n\n\n\n\n\n\nacc\n\n\n\n\n\n\n\n\n\n\n\n\n always produce best models variety \nmeasures using hybrid weighted random forest\nalgorithm\nacknowledgements\n research supported part nsfc \ngrant shenzhen new industry development fund grant nocxba\nreferences\n breiman l random forests machine learning\n \n ho t random subspace method constructing decision forests ieee transactions pattern\nanalysis machine intelligence \n quinlan j c programs machine\nlearning morgan kaufmann\n breiman l classification regression trees\nchapman hall crc\n breiman l bagging predictors\nmachine\nlearning \n ho t random decision forests proceedings\n third international conference document\nanalysis recognition pp ieee\n dietterich t experimental comparison \nthree methods constructing ensembles decision\ntrees bagging boosting randomization machine\nlearning \n\nmac\n\n\n\n\n\n\n\n\n\n\n\nmac\n\n\n\n\n\n\n\n\n\n\n\n\nre\nmic\n\n\n\n\n\n\n\n\n\n\n\nlas\nacc\nmic\n\n\n \n \n\n\n \n \n\n\n \n \n \n \nacc\n\n\n\n\n\n\n\n\n\n\n\n\ntr\nmic\n\n\n\n\n\n\n\n\n\n\n\nlas\nmac\nacc\nmic\n \n \n\n\n\n\n \n\n \n\n\n\n \n \n \n \n\n\n \nmac\n\n\n\n\n\n\n\n\n\n\n\n\nacc\n\n\n\n\n\n\n\n\n\n\n\n\nmac\n\n\n\n\n\n\n\n\n\n\n\nmac\n\n\n\n\n\n\n\n\n\n\n\n\n banfield r hall l bowyer k kegelmeyer w\n comparison decision tree ensemble creation\ntechniques ieee transactions pattern analysis\n machine intelligence \n\n robniksikonja\nm improving random forests\nproceedings th european conference \nmachine learning pp springer\n ho t c decision forests proceedings \n fourteenth international conference pattern\nrecognition pp ieee\n dietterrich t machine learning research four\ncurrent direction artificial intelligence magzine \n\n amaratunga d cabrera j lee y \nenriched random forests bioinformatics \n\n ye y li h deng x huang j \nfeature weighting random forest detection hidden\nweb search interfaces journal computational\nlinguistics chinese language processing \n\n xu b huang j williams g wang q \nye y classifying highdimensional data\n random forests built small subspaces\ninternational journal data warehousing \nmining \n xu b huang j williams g li j ye y\n hybrid random forests advantages mixed\ntrees classifying text data proceedings th\npacificasia conference knowledge discovery \ndata mining springer\n\n computer journal vol \n\n \n\n\n\nhybrid weighted random forests classifying highdimensional data\n biggs d de ville b suen e method\n choosing multiway partitions classification \ndecision trees journal applied statistics \n ture m kurt turhan kurum ozdamar\nk comparing classification techniques \npredicting essential hypertension expert systems \napplications \n begum n ma f ren f automatic text summarization using support vector machine\ninternational journal innovative computing information control \n chen j huang h tian s qu y \nfeature selection text classification naive\nbayes expert systems applications \n\n tan s neighborweighted knearest neighbor\n unbalanced text corpus\nexpert systems \napplications \n pearson k theory contingency \n relation association normal correlation\ncambridge university press\n yang y liu x reexamination \ntext categorization methods proceedings th\ninternational conference research development\n information retrieval pp acm\n han e karypis g centroidbased\ndocument classification analysis experimental\nresults proceedings th european conference \nprinciples data mining knowledge discovery\npp springer\n trec\n\ntext\nretrieval\nconference\nhttp trecnistgov\n lewis\nd\n\nreuters\ntext\ncategorization\ntest\ncollection\ndistribution\n\nhttp wwwresearchattcom lewis\n han e boley d gini m gross r hastings\nk karypis g kumar v mobasher b \nmoore j webace web agent document\ncategorization exploration proceedings nd\ninternational conference autonomous agents pp\n acm\n mccallum nigam k comparison \nevent models naive bayes text classification aaai workshop learning text categorization pp \n\n witten frank e hall m data mining\npractical machine learning tools techniques\nmorgan kaufmann\n\n computer journal vol \n\n \n\n\n\n\nStop words are common words found in a language. Words like for, very, and, of, are, etc, are common stop words. Notice they have been removed from the above text.
We can list the stop words:
length(stopwords("english"))## [1] 174stopwords("english")## [1] "i" "me" "my" "myself" "we"
## [6] "our" "ours" "ourselves" "you" "your"
## [11] "yours" "yourself" "yourselves" "he" "him"
## [16] "his" "himself" "she" "her" "hers"
## [21] "herself" "it" "its" "itself" "they"
## [26] "them" "their" "theirs" "themselves" "what"
## [31] "which" "who" "whom" "this" "that"
## [36] "these" "those" "am" "is" "are"
## [41] "was" "were" "be" "been" "being"
## [46] "have" "has" "had" "having" "do"
## [51] "does" "did" "doing" "would" "should"
## [56] "could" "ought" "i'm" "you're" "he's"
## [61] "she's" "it's" "we're" "they're" "i've"
## [66] "you've" "we've" "they've" "i'd" "you'd"
## [71] "he'd" "she'd" "we'd" "they'd" "i'll"
## [76] "you'll" "he'll" "she'll" "we'll" "they'll"
## [81] "isn't" "aren't" "wasn't" "weren't" "hasn't"
## [86] "haven't" "hadn't" "doesn't" "don't" "didn't"
## [91] "won't" "wouldn't" "shan't" "shouldn't" "can't"
## [96] "cannot" "couldn't" "mustn't" "let's" "that's"
## [101] "who's" "what's" "here's" "there's" "when's"
## [106] "where's" "why's" "how's" "a" "an"
## [111] "the" "and" "but" "if" "or"
## [116] "because" "as" "until" "while" "of"
## [121] "at" "by" "for" "with" "about"
## [126] "against" "between" "into" "through" "during"
## [131] "before" "after" "above" "below" "to"
## [136] "from" "up" "down" "in" "out"
## [141] "on" "off" "over" "under" "again"
## [146] "further" "then" "once" "here" "there"
## [151] "when" "where" "why" "how" "all"
## [156] "any" "both" "each" "few" "more"
## [161] "most" "other" "some" "such" "no"
## [166] "nor" "not" "only" "own" "same"
## [171] "so" "than" "too" "very"Your donation will support ongoing development and give you access to the PDF version of this book. Desktop Survival Guides include Data Science, GNU/Linux, and MLHub. Books available on Amazon include Data Mining with Rattle and Essentials of Data Science. Popular open source software includes rattle, wajig, and mlhub. Hosted by Togaware, a pioneer of free and open source software since 1984.
Copyright © 1995-2021 Graham.Williams@togaware.com Creative Commons Attribution-ShareAlike 4.0.