Difference between revisions of "Structured SVMs"

From Cohen Courses
Jump to navigationJump to search
 
(7 intermediate revisions by the same user not shown)
Line 1: Line 1:
Being edited by Rui Correia
+
== The Method and When to Use it==
 
 
== The Method and When to Use it ==
 
  
 
Structured (or Structural) Support Vector Machines (SSVM), as the name states, is a machine learning model that generalizes the [[Support Vector Machines| Support Vector Machine (SVM)]] classifier, allowing training a classifier for structured output.
 
Structured (or Structural) Support Vector Machines (SSVM), as the name states, is a machine learning model that generalizes the [[Support Vector Machines| Support Vector Machine (SVM)]] classifier, allowing training a classifier for structured output.
  
In general, SSVMs perform supervised learning by approximating a mapping <math>h</math>
+
In general, SSVMs perform supervised learning by approximating a mapping <math>f</math>
  
 
<math>
 
<math>
h: X \rightarrow Y
+
f: X \rightarrow Y
 
</math>
 
</math>
  
Line 26: Line 24:
 
</math>
 
</math>
  
where <math>\Delta</math> is an arbitrary function, which measures the distance between to labels and <math>\Psi</math> is a function on samples and labels, which extracts feature vectors.
+
where <math>\Delta</math> is an arbitrary function, which measures the distance between to labels and <math>\Psi</math> is a function on samples and labels, which extracts feature vectors. These two function are to be defined according to the problem at hand
  
  
Since the regularized risk function above is non-differentiable, it is often reformulated in terms of a quadratic program by introducing one slack variables <math>\xi_i</math> for each sample, each representing the value of the maximum. The standard structured SVM primal formulation is given as follows.
+
Since the equation above is non-differentiable, one can reformulate it introducing slack variables, <math>\xi_n</math>, representing the value of the maximum. Using this approach the SSVM comes as:
  
Slightly
 different 
version 
of 
the
 loss 
function:

+
<math>
 +
\min_{w,\xi} ||w||^2 + C \sum^l_{n=1} \xi_n
 +
</math>
  
<math>
+
      <math>
\min_i \frac{C}{2} ||w||{^{2}}{_{2}} + \sum^N_{i=1} \xi_i
+
s.t.  w'\Psi(x_n,y_n) - w'\Psi(x_n,y) + \xi_n \ge \Delta (y_n,y), n= 1,...,l \forall y \in Y
 
</math>
 
</math>
  
 
       <math>
 
       <math>
s.t. \forall i, \forall y, w^T g (x_i,y) \ge +1 - \frac {\xi_i}{cost(y, y_i)}
+
    \xi_n \ge 0, n= 1,...,l
 +
</math>
 +
 
 +
== Testing ==
 +
Given a sample, <math>x \in X </math> and a mapping <math>f: X \rightarrow Y</math> one can obtain the correspondent lable. The mapping is defined as
 +
 
 +
 
 +
<math>
 +
f(x) = \arg\max_{y \in Y} w'\Psi(x,y)
 
</math>
 
</math>
 +
 +
where <math>w</math> is the vector during the training phase.
 +
 +
The inference problem of solving for this maximizer over the label space is dependent on the structure of the function <math>\Psi</math> that varies according to the problem.
  
 
== Related Papers ==
 
== Related Papers ==
*[http://svmlight.joachims.org/svm_struct.html| I. Tsochantaridis, T. Hofmann, T. Joachims, and Y. Altun. Support Vector Learning for Interdependent and Structured Output Spaces, ICML, 2004.]
+
*[[Taskar, B. et al, NIPS 2003| Taskar 
et 
al.
 (2003)]]
* Optimization Algorithms
+
*[http://svmlight.joachims.org/svm_struct.html| Tsochantaridis et al. (2004)]
**Taskar 
et 
al.
(2003):

 SMO
 based 
on 
factored 
dual

+
*[[Taskar et al. 2004. Max-margin Parsing| Taskar et al. (2004)]]
**Bartlet
 et 
al. 
(2004) 
and 
Collins 
et 
al. 
(2008):

 exponentiated 
gradient

+
*[[Tsochantaridis et al, JMLR 2005| Tsochantaridis 
et 
al. 
(2005)]]
**Tsochantaridis 
et 
al. 
(2005): 

cutting 
planes 
(based 
on
 dual)

 
**Taskar
 et 
al. 
(2005):

 dual 
extragradient

 
**Ratliff
 et
 al.(2006):

 (stochastic) 
subgradient
 descent

 
**Crammer
 et 
al. 
(2006):

 online
“passive‐aggressive”
 algorithms
 

Latest revision as of 17:38, 2 November 2011

The Method and When to Use it

Structured (or Structural) Support Vector Machines (SSVM), as the name states, is a machine learning model that generalizes the Support Vector Machine (SVM) classifier, allowing training a classifier for structured output.

In general, SSVMs perform supervised learning by approximating a mapping

where is a set of labeled training examples and is a complex structured object, like trees, sequences, or sets, instead of simple univariate predictions (as in the SVM case).

Thus, training a SSVM classifier consists of showing pairs of correct sample and output label pairs, that are used for training, allowing to predict for new sample instances the corresponding output label

In NLP one can fing a great variety of problems that rely on complex outputs, such as parsing and Markov Models for part-of-speech tagging.

Training

While on training, for a set of samples and labels , the SSVM minimizes the risk function:


where is an arbitrary function, which measures the distance between to labels and is a function on samples and labels, which extracts feature vectors. These two function are to be defined according to the problem at hand


Since the equation above is non-differentiable, one can reformulate it introducing slack variables, , representing the value of the maximum. Using this approach the SSVM comes as:

      
      

Testing

Given a sample, and a mapping one can obtain the correspondent lable. The mapping is defined as


where is the vector during the training phase.

The inference problem of solving for this maximizer over the label space is dependent on the structure of the function that varies according to the problem.

Related Papers