（5A文）迁移学习算法研究课件.ppt

1、【5A5A文文】迁移学习算法研究迁移学习算法研究TrainingDataOccPalm LinesDragonStarFortune?ProflongTgoodLawyershortFbadPhD StubrokenTgoodDoclongFbadClassifierUnseen Data(,long, T)good!What if2022-6-92传统监督机器学习传统监督机器学习(1/2)(1/2)from Prof. Qiang Yang2022-6-93l传统监督学习同源、独立同分布同源、独立同分布两个基两个基本假设本假设标注足够多的训练样本标注足够多的训练样本在实际应用中在实际应用中通

2、常不能满足！通常不能满足！训练集测试集分类器训练集测试集分类器2022-6-94l实际应用学习场景HP 新闻新闻Lenovo 新闻新闻不同源、分布不一致不同源、分布不一致人工标记训练样本，费人工标记训练样本，费时耗力时耗力迁移迁移学习学习 运用已有的知识对运用已有的知识对不同但相关领域不同但相关领域问题问题进行求解的一种新的机器学习方法进行求解的一种新的机器学习方法 放宽了传统机器学习的两个基本假设放宽了传统机器学习的两个基本假设2022-6-95l迁移学习场景无处不在迁移迁移知识知识迁移迁移知识知识图像分类图像分类HP 新闻新闻Lenovo 新闻新闻新闻网页分类新闻网页分类异构特征空间202

3、2-6-96The apple is the pomaceous fruit of the apple tree, species Malus domestica in the rose family Rosaceae .Banana is the common name for a type of fruit and also the herbaceous plants of the genus Musa which produce this commonly eaten fruit .Training: TextFuture: ImagesApplesBananasfrom Prof. Q

4、iang YangXin Jin, Fuzhen Zhuang, Sinno Jialin Pan, Changying Du, Ping Luo, Qing He: Heterogeneous Multi-task Semantic Feature Learning for Classification. CIKM 2015 : 1847-1850. Test Test Training TrainingClassifierClassifier72.65%DVDElectronicsElectronics84.60%ElectronicsDrop!2022-6-97from Prof. Qi

5、ang Yang2022-6-98DVDElectronicsBookKitchenClothesVideo gameFruitHotelTeaImpractical!from Prof. Qiang YangpConcept Learning for Transfer Learning Concept Learning based on Non-negative Matrix Tri-factorization for Transfer Learning Concept Learning based on Probabilistic Latent Semantic Analysis for

6、Transfer LearningpTransfer Learning using Auto-encodersTransfer Learning from Multiple Sources with Autoencoder RegularizationSupervised Representation Learning: Transfer Learning with Deep Auto-encoders2022-6-99Concept Learning based on Non-negative Matrix Tri-factorization for Transfer Learning202

7、2-6-9Concept Learning for Transfer Learning102022-6-9Concept Learning for Transfer Learning11 Many traditional learning techniques work well only under the assumption: Training and test data follow the same distribution Training (labeled)ClassifierTest (unlabeled)From different companiesEnterprise N

8、ews Classification: including the classes“Product Announcement”, “Business scandal”, “Acquisition”, Product announcement: HPs just-released LaserJet Pro P1100 printer and the LaserJet Pro M1130 and M1210 multifunction printers, price performance .Announcement for Lenovo ThinkPad ThinkCentre price $1

9、50 off Lenovo K300 desktop using coupon code . Lenovo ThinkPad ThinkCentre price $200 off Lenovo IdeaPad U450p laptop using. .their performanceHP newsLenovo newsDifferent distributionFail !2022-6-9Concept Learning for Transfer Learning12 Example AnalysisProduct announcement: HPs just-released LaserJ

10、et Pro P1100 printer and the LaserJet Pro M1130 and M1210 multifunction printers, price performance .Announcement for Lenovo ThinkPad ThinkCentre price $150 off Lenovo K300 desktop using coupon code . Lenovo ThinkPad ThinkCentre price $200 off Lenovo IdeaPad U450p laptop using. .their performanceHP

11、newsLenovo newsProductword conceptLaserJet, printer, price, performance ThinkPad, ThinkCentre, price, performance RelatedProductannouncementdocument class:Share some common words: announcement, price, performance indicate2022-6-9Concept Learning for Transfer Learning13 Example Analysis:HPLaserJet, p

12、rinter, price, performance et al.LenovoThinkpad, Thinkcentre, price, performance et al.The words expressing the same word concept are domain-dependent ProductProductannouncementword conceptindicatesThe association between word concepts and document classes is domain-independent 2022-6-9Concept Learn

13、ing for Transfer Learning14 Further observations:Different domains may use same key words to express the same concept (denoted as identical concept)Different domains may also use different key words to express the same concept (denoted as alike concept)Different domains may also have their own disti

14、nct concepts (denoted as distinct concept) The identical and alike concepts are used as the shared concepts for knowledge transfer We try to model these three kinds of concepts simultaneously for transfer learning text classification2022-6-9Concept Learning for Transfer Learning15 Basic formula of m

15、atrix tri-factorization: where the input X is the word-document co-occurrence matrix denotes concept information, may vary in different domainsF denotes the document classification information indeed is the association between word concepts and document classes, may retain stable cross domainsGS2022

16、-6-9Concept Learning for Transfer Learning16lSketch map of MTrickSource domain Xs FsGsFtGtTargetdomain XtSKnowledge TransferlConsidering the alike conceptslOptimization problem for MTrick2022-6-9Concept Learning for Transfer Learning17G0 is the supervision informationthe association S is shared as b

17、ridge to transfer knowledgelDual Transfer Learning (Long et al., SDM 2012), considering identical and alike concepts2022-6-9Concept Learning for Transfer Learning18lFurther divide the word concepts into three kinds:F1, identical concepts; F2, alike concepts; F3, distinct concepts Input: s source dom

18、ain Xr(1rs) with label information, t target domain Xr (s+1rs+t) We propose Triplex Transfer Learning framework based on matrix tri-factorization (TriTL for short)F1, S1 and S2 are shared as the bridge for knowledge transfer across domainsThe supervision information is integrated by Gr (1rs) in sour

19、ce domainsTriTLTriTL (2 (2/5/5) )lOptimization Problem2022-6-9Concept Learning for Transfer Learning19TriTLTriTL (3 (3/5/5) )lWe develop an alternatively iterative algorithm to derive the solution and theoretically analyze its convergence 2022-6-9Concept Learning for Transfer Learning20TriTLTriTL (4

20、 (4/5/5) )lClassification on target domainsWhen 1rs, Gr contains the label information, so we remain it unchanged during the iterations when xi belongs to class j, then Gr(i,j)=1, else Gr(i,j)=0After the iteration, we obtain the output Gr (s+1rs+t), then we can perform classification according to Gr

21、2022-6-9Concept Learning for Transfer Learning21TriTLTriTL (5 (5/5/5) )lAnalysis of Algorithm ConvergenceAccording to the methodology of convergence analysis in the two works Lee et al., NIPS01 and Ding et al., KDD06, the following theorem holds.Theorem (Convergence): After each round of calculating

22、 the iterative formulas, the objective function in the optimization problem will converge monotonically.2022-6-9Concept Learning for Transfer Learning222022-6-9Concept Learning for Transfer Learning23rec.autosrec.motorcyclesrec.baseballrec.hockeysci.cryptsic.electronicssci.medsci.spacecomp.graphicsc

23、omp.sys.ibm.pc.hardwarecomp.sys.mac.hardwarecomp.windows.xtalk.politics.misctalk.politics.gunstalk.politics.mideasttalk.religion.miscrecscicomptalkData Preparation (1Data Preparation (1/3/3) )l20NewsgroupsFour top categories, each top category contains four sub-categorieslSentiment Classification, f

24、our domains: books, dvd, electronics, kitchen Randomly select two domains as sources, and the rest as targets, then 6 problems can be constructed2022-6-9Concept Learning for Transfer Learning24rec.autosrec.motorcyclesrec.baseballrec.hockeysci.cryptsic.electronicssci.medsci.spacerec +sci -baseball cr

25、ypy Source domainautos spaceT a r g e t domainlFor the classification problem with one source domain and one target domain, we can construct 144 ( ) problems2244PPData Preparation (2Data Preparation (2/3/3) )lConstruct classification tasks (Traditional TL)2022-6-9Concept Learning for Transfer Learni

26、ng25lConstruct new transfer learning problems rec.autosrec.motorcyclesrec.baseballrec.hockeysci.cryptsic.electronicssci.medsci.spacerec +sci -baseball crypy autos spacecomp.graphicscomp.sys.ibm.pc.hardwarecomp.sys.mac.hardwarecomp.windows.xtalk.politics.misctalk.politics.gunstalk.politics.mideasttal

27、k.religion.misccomptalkautos graphics14483384 !More distinct concepts may exist!Data Preparation (3Data Preparation (3/3/3) )Source domainT a r g e t domain2022-6-9Concept Learning for Transfer Learning26Compared AlgorithmsCompared AlgorithmslTraditional learning AlgorithmsSupervised Learning: Logis

28、tic Regression (LR) David et al., 00Support Vector Machine (SVM) Joachims, ICML99Semi-supervised Learning: TSVM Joachims, ICML99lTransfer learning Methods: CoCC Dai et al., KDD07, DTL Long et al., SDM12lClassification accuracy is used as the evaluation measure 2022-6-9Concept Learning for Transfer L

29、earning27Experimental Results (1Experimental Results (1/3/3) )lSort the problems with the accuracy of LRDegree of transfer difficultyeasierlGenerally, the lower of accuracy of LR can indicate the harder to transfer, while the higher ones indicate the easier to transferharder2022-6-9Concept Learning

30、for Transfer Learning28Experimental Results (2Experimental Results (2/3/3) )lComparisons among TriTL, DTL, MTrick, CoCC, TSVM, SVM and LR on data set rec vs. sci (144 problems)TriTL can perform well even the accuracy of LR is lower than 65%2022-6-9Concept Learning for Transfer Learning29Experimental

31、 Results (3Experimental Results (3/3/3) )lResults on new transfer learning problems, we only select the problems, whose accuracies of LR are between (50%, 55% (Only slightly better than random classification, thus they might be much more difficult).lWe obtain 65 problems lTriTL also outperforms all

32、the baselines2022-6-9Concept Learning for Transfer Learning30Explicitly define three kinds of word concepts, i.e., identical concept, alike concept and distinct conceptPropose a general transfer learning framework based on nonnegative matrix tri-factorization, which simultaneously model the three ki

33、nds of concepts (TriTL) Extensive experiments show the effectiveness of the proposed approach, especially when the distinct concepts may existConcept Learning based on Probabilistic Latent Semantic Analysis for Transfer Learning2022-6-9Concept Learning for Transfer Learning312022-6-9Concept Learning

34、 for Transfer Learning32MotivationMotivationProduct announcement: HPs just-released LaserJet Pro P1100 printer and the LaserJet Pro M1130 and M1210 multifunction printers, price performance .Announcement for Lenovo ThinkPad ThinkCentre price $150 off Lenovo K300 desktop using coupon code . Lenovo Th

35、inkPad ThinkCentre price $200 off Lenovo IdeaPad U450p laptop using. .their performanceHP newsLenovo newsProductword conceptLaserJet, printer, price, performance ThinkPad, ThinkCentre, price, performance RelatedProductannouncementdocument class:Share some common words: announcement, price, performan

36、ce indicatelRetrospect the example2022-6-9Concept Learning for Transfer Learning33lSome notationsddocumentydocument classzword conceptlSome definitionse.g., p(price|Product), p(LaserJet|Product,)wwordrdomaine.g, p(Product|Product announcement)Preliminary Knowledge (1Preliminary Knowledge (1/3/3) )20

37、22-6-9Concept Learning for Transfer Learning34Preliminary Knowledge (2Preliminary Knowledge (2/3/3) )ProductLaserJet, printer, announcement, price, ThinkPad, ThinkCentre, announcement, price Productannouncementp(w|z,r1)p(w|z,r2)p(z|y) p(w|z,r1) p(w|z,r2) E.g., p(LaserJet|Product, HP) p(LaserJet|Prod

38、uct, Lenovo) p(z|y,r1) = p(z|y,r2)E.g., p(Product|Product annoucement, HP) = p(Product|Product annoucement, Lenovo)lAlike concept2022-6-9Concept Learning for Transfer Learning35lDual PLSA (D-PLSA)lJoint probability over all variables p(w,d) = p(w|z) p(z|y) p(d|y) p(y)lGiven data domain X, the proble

39、m of maximum log likelihood islog p(X;) = log z p(Z,X;) includes all the parameters p(w|z), p(z|y), p(d|y), p(y). Z denotes all the latent variablesPreliminary Knowledge (3/3/3)lThe proposed transfer learning algorithm based on D-PLSA, denoted as HIDC 2022-6-9Concept Learning for Transfer Learning36

40、lIdentical conceptp(w|za)p(za|y)lAlike conceptThe extension and intension are domain independentp(w|zb,r)p(zb|y)HIDC (1/3/3)The extension is domain dependent, while the intension is domain independent2022-6-9Concept Learning for Transfer Learning37lDistinct conceptp(w|zc,r)p(zc|y,r)lThe joint probab

41、ilities of these three graphical modelsHIDC (2/3/3)The extension and intension are both domain dependent2022-6-9Concept Learning for Transfer Learning38lGiven s+t data domains X = X1, Xs, Xs+1, Xs+t, without loss of generality, the first s domains are source domains, and the left t domains are targe

42、t domainslConsider the three kinds of concepts:lThe Log likelihood function islog p(X;) = log z p(Z,X;) includes all parameters p(w|za), p(w|zb,r), p(w|zc,r), p(za|y), p(zb|y), p(zc|y,r), p(d|y,r), p(y|r), p(r).HIDC (3/3/3)2022-6-9Concept Learning for Transfer Learning39lUse the EM algorithm to deri

43、ve the solutionslE Step:Model Solution (1/4/4)2022-6-9Concept Learning for Transfer Learning40lM Step：Model Solution (2/4/4)2022-6-9Concept Learning for Transfer Learning41lSemi-supervised EM algorithm: when r is from source domains, the labeled information p(d|y,r) is known and p(y|r) can be infere

44、d p(d|y,r) = 1/ny,r, if d belongs y in domain r, ny,r is the number of documents in class y in domain r, else p(d|y,c) = 0 p(y|r) = ny,r / nr , nr is the number of documents in domain r when r is from source domains, p(d|y,r) and p(y|r) keep unchanged during the iterations, which supervise the optim

45、izing processModel Solution (3/4/4)2022-6-9Concept Learning for Transfer Learning42lClassification for target domains After we obtain the final solutions of p(w|za), p(w|zb,r), p(w|zc,r), p(za|y), p(zb|y), p(zc|y,r), p(d|y,r), p(y|r), p(r) We can compute the conditional probabilities: Then the final

46、 prediction isDuring the iterations, all domains share p(w|za), p(za|y), p(zb|y), which act as the bridge for knowledge transferModel Solution (4/4/4)2022-6-9Concept Learning for Transfer Learning43BaselineslCompared AlgorithmsSupervised Learning: pLogistic Regression (LG) David et al., 00pSupport V

47、ector Machine (SVM) Joachims, ICML99Semi-supervised Learning: pTSVM Joachims, ICML99Transfer Learning: pCoCC Dai et al., KDD07 pCD-PLSA Zhuang et al., CIKM10 pDTL Long et al., SDM12lOur MethodspHIDClMeasure: classification accuracy2022-6-9Concept Learning for Transfer Learning44lResults on new trans

48、fer learning problems, we select the problems, whose accuracies of LR are higher than 50%, then 334 problems are obtainedExperimental Results (1/5/5)2022-6-9Concept Learning for Transfer Learning45lResults on new transfer learning problems, we select the problems, whose accuracies of LR are higher t

49、han 50%, then 334 problems are obtainedExperimental Results (2/5/5)2022-6-9Concept Learning for Transfer Learning46Experimental Results (3/5/5)2022-6-9Concept Learning for Transfer Learning47Source domain: S (rec.autos, sci.space)，Target domain: T(rec.sport.hockey, talk.politics.mideast)STSTDistinct

50、 conceptSTAlike conceptExperimental Results (4/5/5)2022-6-9Concept Learning for Transfer Learning48Experimental Results (5/5/5)lIndeed, the proposed probabilistic method HIDC is also better than TriTLlThis may due to the reason that there is more clearer probabilistic explanation of HIDC p1(z, y) =