Computing Reviews
Today's Issue Hot Topics Search Browse Recommended My Account Log In
Review Help
Search
A filter attribute selection method based on local reliable information
Martín R., Aler R., Galván I. Applied Intelligence48 (1):35-45,2018.Type:Article
Date Reviewed: Apr 25 2018

In classification algorithms, the core problem is selecting the right attributes and assigning them the right weight for each item being processed, in order to achieve reliable results; all the more so if machine learning is involved. At present, of course, several algorithms are routinely in use in many branches of industry and research with the full satisfaction of their respective users; nevertheless, science and research always look to the future in order to improve existing solutions. This paper is an example of this line of research.

It starts by examining current algorithms. Most evaluate single attributes, however numerous, of items under scrutiny; they use two techniques, filtering and wrapping. Filtering means attribute selection according to some general characteristics of data; wrapping means attribute selection according to some given machine learning algorithm. Generally, wrapping is more accurate, but becomes computationally demanding when the number of features is large; moreover, overfitting is a risk if very detailed sets of rules are used. In these cases, filtering becomes faster; to overcome these problems, hybrid systems, filter-then-wrap, are often used. Often, all of these methods present high noise levels caused by too few items fitting the desired parameters, class imbalance (wildly varying the number of items in each class), or simply class overlap.

To overcome these shortcomings, the authors propose a solution evaluating not only single attributes of items, but also the interactions among those attributes. Items are first classified by placing them into several attribute grids defined by the user, where grid axes are the attributes to be evaluated. Grids with one, two, or three axes can be built, but no more because additional axes do not give better results. The final weight for each item is given by the sum of weights in each cell of the grid. This is the filter part of the solution. Then, the information content for each cell is evaluated via an algorithm, which constitutes the wrap part of the solution.

The authors finally test this solution against the most widespread algorithms currently in use; in most cases, their algorithm comes to more reliable results (it never gives worse results). A nice feature of the paper is the way results are displayed, with a series of pleasant and intuitive diagrams. Moreover the algorithm is revealed to be computationally efficient by directly using parallelism and integer arithmetic instead of computationally heavy conditional structures typical of programming languages.

All of this is explained in a terse and linear paper that for each topic comes straight to the point without unnecessary frills. It certainly makes for interesting reading, but will be best appreciated with a solid scientific background.

Reviewer:  Andrea Paramithiotti Review #: CR145997 (1807-0404)
Bookmark and Share
  Featured Reviewer  
 
Feature Evaluation And Selection (I.5.2 ... )
 
 
Filtering (I.4.3 ... )
 
 
Learning (I.2.6 )
 
Would you recommend this review?
yes
no
Other reviews under "Feature Evaluation And Selection": Date
Labeled point pattern matching by Delaunay triangulation and maximal cliques
Ogawa H. Pattern Recognition 19(1): 35-40, 1986. Type: Article
Feb 1 1988
Features selection and ‘possibility theory’
Di Gesù V., Maccarone M. Pattern Recognition 19(1): 63-72, 1986. Type: Article
Dec 1 1987
An analytic-to-holistic approach for face recognition based on a single frontal view
Lam K., Yan H. IEEE Transactions on Pattern Analysis and Machine Intelligence 20(7): 673-686, 1998. Type: Article
Oct 1 1998
more...

E-Mail This Printer-Friendly
Send Your Comments
Contact Us
Reproduction in whole or in part without permission is prohibited.   Copyright 1999-2024 ThinkLoud®
Terms of Use
| Privacy Policy