Feature evaluation: Unsupervised model and supervised model

Unsupervised model

Here,

• Variance can be used as evaluation criteria.
• The maximum variance ranked first.
• Descending order sorting of variances.

Fig: Unsupervised feature selection model.

Rank the features based on variance and chose top "d" features.

This is called unsupervised feature selection or elimination model. 

Supervised model: 

Fig: Supervised feature selection model.

Equation: Mean.

Fig: Mean of individual classes.

Here,

Fig: ICV < BCV.

• The minimum variance ranked first.
• Ascending order sorting of variance.
  
  Compute:
  
  Intra class variance (ICV)/within class variance:
  
  
  
  Inter class variance/between class variance(BCV):
  
  
  Ration from BCV to ICV is high.

  

  Fig: Ratio from BCV to ICV.

  Chose the one feature with the highest rank as first. 
  This is called as Ward's or Fisher's criterion.
  This is only for supervised learning.
  This is used in the Fisher's Linear Discriminant Analysis technique (FLD or LD) and feature transformation.
  Such a discriminating algorithm is called the Fisher's Linear Discrimination algorithm.
  Loosely coupled across the sample of the class. 
  
  

  Example:

  Read the below feature matrix. Calculate the variance sort them accordingly.

Object A, B, C, D, E, and F can be represented by using 8 features.

Table: Feature matrix

Solution:

Table: Solution to the problem stated above.

Ascending sorting of the variances: 

Table: Ascending sorting of variances.

The next topic is filters and wrapper methods. 

Subscribe, to get instant notification of updates.

Follow and don't forget to share.