multipl kernel learn improv mmd effici numer stabl spars learn cross valid framework choos model dataset transfer learn gener heurist dfree featur construct improv accuraci gener heurist dfree featur construct improv accuraci relax transfer class spectral partit abstract exist transfer learn techniqu limit problem knowledg transfer task share set class label paper relax constraint propos spectral base solut aim unveil intrins structur data gener partit target data transfer eigenspac separ sourc data clusteringbas kl diverg propos automat adjust transfer evalu propos model text imag dataset class categori sourc target data explicitli class transfer class show propos approach improv baselin averag accuraci sourc code dataset author univers learn relat distribut adapt graph transduct naiv bay classif uncertain data cross domain distribut adapt kernel map label exampl limit difficult obtain transfer learn emploi knowledg sourc domain improv learn accuraci target domain assumpt made exist approach margin condit probabl directli relat sourc target domain limit applic origin space linear transform solv problem propos adapt kernel approach map margin distribut targetdomain sourc domain data common kernel space util sampl select strategi draw condit probabl domain closer formal show kernel map space differ distribut domain bound predict error propos approach bound experiment result demonstr propos method outperform tradit induct classifi state art boost base transfer algorithm domain includ text categor web page rate achiev higher accuraci approach text categor problem sourc code dataset author latent space domain transfer high dimension overlap distribut transfer knowledg domain challeng due number reason condit margin distribut train data test data ident model train domain directli appli domain low accuraci applic larg featur set text document sequenc data medic data imag data resolut domain featur introduc larg number miss valu consid union featur domain word margin distribut overlap time problem high dimension thousand featur combin high activ transfer domain knowledg graph dbase iter hybrid featur select forward semi dsupervis featur select type dindepend correct sampl select bia structur discoveri dbalanc sampl select bia common problem real world applic train data obtain realist constraint make follow distribut futur test data applic hospit clinic studi common practic build model elig volunt train data appli model entir popul volunt select random train set drawn distribut test set dataset suffer sampl select bia covari shift past year work propos reduc sampl select bia static match distribut train set test set paper explor distribut directli propos high ddimension share nearest neighbor cluster algorithm 