Presentation is loading. Please wait.

Presentation is loading. Please wait.

Semi-supervised Learning Rong Jin. Semi-supervised learning  Label propagation  Transductive learning  Co-training  Active learning.

Similar presentations


Presentation on theme: "Semi-supervised Learning Rong Jin. Semi-supervised learning  Label propagation  Transductive learning  Co-training  Active learning."— Presentation transcript:

1 Semi-supervised Learning Rong Jin

2 Semi-supervised learning  Label propagation  Transductive learning  Co-training  Active learning

3 Label Propagation  A toy problem Each node in the graph is an example  Two examples are labeled  Most examples are unlabeled Compute the similarity between examples S ij Connect examples to their most similar examples  How to predicate labels for unlabeled nodes using this graph? Unlabeled example Two labeled examples w ij

4 Label Propagation  Forward propagation

5 Label Propagation  Forward propagation

6 Label Propagation  Forward propagation How to resolve conflicting cases What label should be given to this node ?

7 Label Propagation  Let S be the similarity matrix S=[S i,j ] nxn  Let D be a diagonal matrix where D i =  i  j S i,j  Compute normalized similarity matrix S’ S’=D -1/2 SD -1/2  Let Y be the initial assignment of class labels Y i = 1 when the i-th node is assigned to the positive class Y i = -1 when the i-th node is assigned to the negative class Y i = 0 when the I-th node is not initially labeled  Let F be the predicted class labels The i-th node is assigned to the positive class if F i >0 The i-th node is assigned to the negative class if F i < 0

8 Label Propagation  Let S be the similarity matrix S=[S i,j ] nxn  Let D be a diagonal matrix where D i =  i  j S i,j  Compute normalized similarity matrix S’ S’=D -1/2 SD -1/2  Let Y be the initial assignment of class labels Y i = 1 when the i-th node is assigned to the positive class Y i = -1 when the i-th node is assigned to the negative class Y i = 0 when the i-th node is not initially labeled  Let F be the predicted class labels The i-th node is assigned to the positive class if F i >0 The i-th node is assigned to the negative class if F i < 0

9 Label Propagation  One iteration F = Y +  S’Y = (I +  S’)Y  weights the propagation values  Two iteration F =Y +  S’Y +  2 S’ 2 Y = (I +  S’ +  2 S’ 2 )Y  How about the infinite iteration F = (  n=0 1  n S’ n )Y = (I -  S’) -1 Y  Any problems with such an approach?

10 Label Consistency Problem  Predicted vector F may not be consistent with the initially assigned class labels Y

11 Energy Minimization  Using the same notation S i,j : similarity between the I-th node and j-th node Y: initially assigned class labels F: predicted class labels  Energy: E(F) =  i,j S i,j (F i – F j ) 2  Goal: find label assignment F that is consistent with labeled examples Y and meanwhile minimizes the energy function E(F)

12 Harmonic Function  E(F) =  i,j S i,j (F i – F j ) 2 = F T (D-S)F  Thus, the minimizer for E(F) should be (D-S)F = 0, and meanwhile F should be consistent with Y.  F T = (F l T, F u T ), Y T = (Y l T, Y u T ) F l = Y l 

13 Optical Character Recognition  Given an image of a digit letter, determine its value 1 2  Create a graph for images of digit letters

14 Optical Character Recognition  #Labeled_Examples+#Unlabeled_Examples = 4000  CMN: label propagation  1NN: for each unlabeled example, using the label of its closest neighbor

15 Spectral Graph Transducer  Problem with harmonic function  Why this could happen ?  The condition (D-S)F = 0 does not hold for constrained cases

16 Spectral Graph Transducer  Problem with harmonic function  Why this could happen ?  The condition (D-S)F = 0 does not hold for constrained cases

17 Spectral Graph Transducer min F F T LF + c (F-Y) T C(F-Y) s.t. F T F=n, F T e = 0  C is the diagonal cost matrix, C i,i = 1 if the i-th node is initially labeled, zero otherwise  Parameter c controls the balance between the consistency requirement and the requirement of energy minimization  Can be solved efficiently through the computation of eigenvector

18 Empirical Studies

19 Problems with Spectral Graph Transducer min F F T LF + c (F-Y) T C(F-Y) s.t. F T F=n, F T e = 0  The obtained solution is different from the desirable one: minimize the energy function and meanwhile is consistent with labeled examples Y  It is difficult to extend the approach to multi-class classification

20 Green’s Function  The problem of minimizing energy and meanwhile being consistent with initially assigned class labels can be formulated into Green’s function problem  Minimizing E(F) = F T LF  LF = 0 Turns out L can be viewed as Laplacian operator in the discrete case LF = 0  r 2 F=0  Thus, our problem is find solution F r 2 F=0, s.t. F = Y for labeled examples We can treat the constraint that F = Y for labeled examples as boundary condition (Von Neumann boundary condition) A standard Green function problem

21 Why Energy Minimization? Final classification results

22 Label Propagation  How the unlabeled data help classification?  Consider a smaller number of unlabeled example

23 Label Propagation  How the unlabeled data help classification?  Consider a smaller number of unlabeled example  Classification results can be very different

24 Cluster Assumption  Cluster assumption Decision boundary should pass low density area  Unlabeled data provide more accurate estimation of local density

25 Cluster Assumption vs. Maximum Margin  Maximum margin classifier (e.g. SVM) denotes +1 denotes -1 w  x+b  Maximum margin  low density around decision boundary  Cluster assumption  Any thought about utilizing the unlabeled data in support vector machine?

26 Transductive SVM  Decision boundary given a small number of labeled examples

27 Transductive SVM  Decision boundary given a small number of labeled examples  How will the decision boundary change given both labeled and unlabeled examples?

28 Transductive SVM  Decision boundary given a small number of labeled examples  Move the decision boundary to place with low local density

29 Transductive SVM  Decision boundary given a small number of labeled examples  Move the decision boundary to place with low local density  Classification results  How to formulate this idea?

30 Transductive SVM: Formulation  Labeled data L:  Unlabeled data D:  Maximum margin principle for mixture of labeled and unlabeled data For each label assignment of unlabeled data, compute its maximum margin Find the label assignment whose maximum margin is maximized

31 Tranductive SVM Different label assignment for unlabeled data  different maximum margin

32 Transductive SVM: Formulation Original SVM Transductive SVM Constraints for unlabeled data A binary variables for label of each example

33 Computational Issue  No longer convex optimization problem. (why?)  How to optimize transductive SVM?  Alternating optimization

34 Alternating Optimization  Step 1: fix y n+1,…, y n+m, learn weights w  Step 2: fix weights w, try to predict y n+1,…, y n+m (How?)

35 Empirical Study with Transductive SVM  10 categories from the Reuter collection  3299 test documents  1000 informative words selected using MI criterion

36 Co-training for Semi-supervised Learning  Consider the task of classifying web pages into two categories: category for students and category for professors  Two aspects of web pages should be considered Content of web pages  “I am currently the second year Ph.D. student …” Hyperlinks  “My advisor is …”  “Students: …”

37 Co-training for Semi-Supervised Learning

38 It is easy to classify the type of this web page based on its content It is easier to classify this web page using hyperlinks

39 Co-training  Two representation for each web page Content representation: (doctoral, student, computer, university…) Hyperlink representation: Inlinks: Prof. Cheng Oulinks: Prof. Cheng

40 Co-training: Classification Scheme 1. Train a content-based classifier using labeled web pages 2. Apply the content-based classifier to classify unlabeled web pages 3. Label the web pages that have been confidently classified 4. Train a hyperlink based classifier using the web pages that are initially labeled and labeled by the classifier 5. Apply the hyperlink-based classifier to classify the unlabeled web pages 6. Label the web pages that have been confidently classified

41 Co-training  Train a content-based classifier

42 Co-training  Train a content-based classifier using labeled examples  Label the unlabeled examples that are confidently classified

43 Co-training  Train a content-based classifier using labeled examples  Label the unlabeled examples that are confidently classified  Train a hyperlink-based classifier Prof. : outlinks to students

44 Co-training  Train a content-based classifier using labeled examples  Label the unlabeled examples that are confidently classified  Train a hyperlink-based classifier Prof. : outlinks to students  Label the unlabeled examples that are confidently classified

45 Co-training  Train a content-based classifier using labeled examples  Label the unlabeled examples that are confidently classified  Train a hyperlink-based classifier Prof. : outlinks to  Label the unlabeled examples that are confidently classified


Download ppt "Semi-supervised Learning Rong Jin. Semi-supervised learning  Label propagation  Transductive learning  Co-training  Active learning."

Similar presentations


Ads by Google