Creating one-class datasets

The first and most important thing for the application of one-class classifiers, is the dataset and its preprocessing. All one-class classifiers require a dataset Matlab object with objects labeled target or outlier. To create a one-class dataset, several functions are supplied: gendatoc, oc_set and target_class. What are the differences between the three?

• gendatoc: this function basically constructs from two Matlab arrays a one-class dataset. When you have two datasets xt and xo available, you can create a one-class dataset using:

    >> xt = randn(40,2);
    >> xo = gendatb([20,0]);
    >> x = gendatoc(xt,xo);
  

  In gendatoc xt or xo do not have to be defined, they can be empty (xt = [] or xo = []). To label a Matlab array as outlier, is therefore every easily done:

    >> xo = 10*randn(25,2);
    >> x = gendatoc([],xo);
  

  If xt or xo is a Prtools dataset, this data is converted back to normal Matlab arrays. That means that the label information in these datasets is lost. All data in xt will be labeled target and all data in xo outlier, without exception.

• oc_set: this function relabels an existing Prtools dataset such that one of the classes becomes target class, and all others become outlier. You have to supply the label of the class that you want to be target class. Assume you generate data from a banana-shaped distribution, and you want to have the class labeled 1 to be target class:

    >> x = gendatb([20,20]);       % 40 objects in 2D
    >> x = oc_set(x,'1')
  

  Now you still have 40 objects, half is labeled target, the other half outlier.

  This function oc_set also accepts several classes to be labeled as target class. When a 10-class problem is loaded, a subset of these classes can be assigned to be target class:

    >> load nist16;   % this is an example of a 10-class dataset,
    >>                % it might not be available everywhere
    >> a
  	2000 by 256 dataset with 10 classes: [200  200  200
  	200  200  200  200  200  200  200]
    >> x = oc_set(a,[1 5 6])  %select three classes
  	Class 0 is used as target class.
  	Class 4 is used as target class.
  	Class 5 is used as target class.
  	 (3 classes as target), 2000 by 256 dataset with 2
  	 classes: [600  1400]
  

  When you don't supply labels, it is assumed that all data is target data:

    >> x = rand(20,2);
    >> x = oc_set(x)
  

  This constructs a dataset, containing 20 target objects in 2D. All objects are now labeled target. When you want to label this data as outlier, you have to supply it as the second argument: x = oc_set([],x).

• target_class: this functions labels one of the classes as target (identical to oc_set) but next removes all other objects from the dataset.

    >> x = gendatb([20,20]);       % 40 objects in 2D
    >> x = target_class(x,'1')     % 20 objects in 2D
  

  Now dataset x just contains 20 target objects. You can achieve the same in this way:

    >> x = gendatb([20,20]);       % 40 objects in 2D
    >> x = oc_set(x,'1');
    >> x = target_class(x)         % 20 objects in 2D
  

  but this is not so efficient.

  In some cases you may need to extract the outlier data. This is obtained as the second output argument from target_class:

    >> [xt,xo] = target_class(x)   % xo contains 20 outlier objects