improv cluster stabil combinatori mrf data weav scale state dof dthe dart data cluster dclass cluster text domain multi dmodal cluster multimedia collect web page cluster heurist search web graph combinatori markov random field abstract combinatori random variabl discret random variabl defin combinatori set power set set paper introduc combinatori markov random field comraf markov random field node combinatori random variabl argu comraf power model unsupervis semi supervis learn put comraf perspect show relationship exist model problemat appli exist infer techniqu graphic model comraf design simpl effici infer algorithm specif comraf base combinatori optim show simpl algorithm consist significantli outperform latent dirichlet alloc lda document cluster task present comraf model semi supervis cluster transfer learn demonstr superior result comparison exist semi supervis scheme constrain optim multi dwai distribut cluster pairwis interact present unsupervis learn scheme simultan cluster variabl type document word author base pairwis interact type observ occurr data scheme multipl cluster system gener aim maxim object function measur multipl pairwis mutual inform cluster variabl implement idea propos algorithm interleav top cluster variabl bottom cluster variabl local optim correct routin focus document cluster present extens empir studi applic scheme real world dataset includ newsgroup ng enron email collect multi distribut cluster mdc algorithm consist significantli outperform previou state art inform theoret cluster algorithm disambigu web appear peopl social network inform person search engin return page person page person care peopl happen multipl peopl relat leverag social network paper present unsupervis framework solv problem base link structur web page agglom conglom doubl cluster cdc applic recent introduc multi distribut cluster method evalu method collect hand label dataset web page retriev googl queri person name appear someon email folder dataset method outperform tradit agglom cluster achiev measur extract social network contact inform email web abstract present end end system extract user social network member contact inform user email inbox system identifi uniqu peopl email find web presenc automat fill field contact address book condit random field type probabilist model suit inform extract task recurs call peopl discov web system build social network multipl degre separ user addition set expertis describ keyword extract person outlin collect statist learn compon enabl system present experiment result real email user present result simpl method learn transfer discuss capabl system addressbook popul expert find social network analysi distribut word cluster word text categor studi approach text categor combin distribut cluster word support vector machin svm classifi word cluster represent comput recent introduc inform bottleneck method gener compact effici represent document combin classif power svm method yield high perform text categor combin svm word cluster represent compar svm base categor simpler bag word bow represent comparison perform dataset dataset newsgroup method base word cluster significantli outperform word base represent term categor accuraci represent effici set reuter webkb word base represent slightli outperform word cluster represent investig potenti reason behavior relat structur differ dataset featur distribut cluster text categor describ power text categor method base combin distribut featur support vector machin svm classifi featur select approach distribut cluster word recent introduc inform bottleneck method gener cient represent document combin classif power support vector machin produc multilabel categor result newsgroup dataset 