1. Feature Selection: Algorithms and Challenges
Joint Work with Yanglan Gang, Hao Wang & Xuegang Hu
Xindong Wu
University of Vermont, USA;
Hefei University of Technology, China
合肥工业大学计算机应用长江学者讲座教授

2. Deduction Induction: My Research Background
1988
Expert Systems
2004
1990
Expert Systems
1995 …

3. Outlines Why feature selection What is feature selection
Components of feature selection
Some research efforts by myself
Challenges in feature selection

4. 1. Why Feature Selection?
High-dimensional data often contain irrelevant or redundant features
reduce the accuracy of data mining algorithms
slow down the mining process
be a problem in storage and retrieval
hard to interpret

5. 2. What Is Feature Selection?
Select the most “relevant” subset of attributes according to some selection criteria.

6. Outlines Why feature selection What is feature selection
Components of feature selection
Some research efforts by myself
Challenges in feature selection

7. Traditional Taxonomy Wrapper approach Filter approach
Features are selected as part of the mining algorithm
Filter approach
Features selected before a mining algorithm,using heuristics based on general characteristics of the data, rather than a learning algorithm to evaluate the merit of feature subsets
Wrapper approach is generally more accurate but also more computationally expensive.

8. Components of Feature Selection
Feature selection is actually a search problem, including four basic components:
an initial subset
one or more selection criteria (＊)
a search strategy (＊)
some given stopping conditions

9. Feature Selection Criteria
Selection criteria generally use “relevance” to estimate the goodness of a selected feature subset in one way or another:
Distance Measure
Information Measure
Inconsistency Measure
Relevance Estimation
Selection Criteria related to Learning Algorithms (wrapper approach)
Some unified framework for relevance has been proposed recently.

10. Search Strategy Exhaustive Search A modified approach: B&B
Every possible subset is evaluated and the best one is chosen
Guarantee the optimal solution
Low efficiency
A modified approach: B&B

11. Search Strategy (2) Heuristic search
Sequential search, including SFS,SFFS,SBS and SBFS
SFS: Start with empty attribute set
Add “best” of attributes
Add “best” of remaining attributes
Repeat until the maximum performance is reached
SBS: Start with the entire attribute set
Remove “worst” of attributes
Repeat until the maximum performance has been reached.

12. Search Strategy (3) Random search It proceeds in two different ways
Inject randomness into classical sequential approaches (simulated annealing, beam search, the genetic algorithm , and random-start hill-climbing)
Generate the next subset randomly
The use of randomness can help to escape local optima in the search space, and the optimality of the selected subset would depend on the available resources.

13. Outlines Why feature selection What is feature selection
Components of feature selection
Some research efforts by myself
Challenges in feature selection

14. RITIO: Rule Induction Two In One
Feature selection using the information gain in a reverse order
Delete features that are lest informative
Results are significant compared to forward selection
[Wu et al 1999, TKDE].

15. Induction as Pre-processing
Use one induction algorithm to select attributes for another induction algorithm
Can be a decision-tree method for rule induction, or vice versa
Accuracy results are not as good as expected
Reason: feature selection normally causes information loss
Details: [Wu 1999, PAKDD].

16. Subspacing with Asysmetric Bagging
When the number of examples is less than the number of attributes
When the number of positive examples is smaller than the number of negative examples
An example: content-based information retrieval
Details: [Tao et al., 2006, TPAMI].

17. Outlines Why feature selection What is feature selection
Components of feature selection
Some research efforts by myself
Challenges in feature selection

18. Challenges in Feature Selection (1)
Dealing with ultra-high dimensional data and feature interactions
Traditional feature selection encounter two major problems when the dimensionality runs into tens or hundreds of thousands:
curse of dimensionality
the relative shortage of instances.
1 curse of dimensionality
As most existing feature selection algorithms have quadratic or higher time complexity about N, it is difficult to scale up with high dimensionality.
2 the relative shortage of instances.
That is, the dimensionality N can sometimes greatly exceed the number of instances I

19. Challenges in Feature Selection (2)
Dealing with active instances (Liu et al., 2005)
When the dataset is huge, feature selection performed on the whole dataset is inefficient,
so instance selection is necessary:
Random sampling (pure random sampling without exploiting any data characteristics)
Active feature selection (selective sampling using data characteristics achieves better or equally good results with a significantly smaller number of instances).
Random sampling (pure random sampling without exploiting any data characteristic)
Active feature selection (selective sampling by using data characteristics achieves better or equally good results with a significantly smaller number of instances)

20. Challenges in Feature Selection (3)
Dealing with new data types (Liu et al., 2005)
traditional data type: an N*M data matrix
Due to the growth of computer and Internet/Web techniques, new data types are emerging:
text-based data (e.g., s, online news, newsgroups)
semistructure data (e.g., HTML, XML)
data streams.

21. Challenges in Feature Selection (4)
Unsupervised feature selection
Feature selection vs classification: almost every classification algorithm
Subspace method with the curse of dimensionality in classification
Subspace clustering.

22. Challenges in Feature Selection (5)
Dealing with predictive-but-unpredictable attributes in noisy data
Attribute noise is difficult to process, and removing noisy instances is dangerous
Predictive attributes: essential to classification
Unpredictable attributes: cannot be predicted by the class and other attributes
Noise identification, cleansing, and measurement need special attention [Yang et al., 2004]

23. Challenges in Feature Selection (6)
Deal with inconsistent and redundant features
Redundancy can indicate reliability
Inconsistency can also indicate a problem for handling
Researchers in Rough Set Theory: What is the purpose of feature selection?
Can you really demonstrate the usefulness of reduction, in data mining accuracy, or what?
Removing attributes can well result in information loss
When the data is very noisy, removals can cause a very different data distribution
Discretization can possibly bring new issues.

24. Concluding Remarks
Feature selection is and will remain an important issue in data mining, machine learning, and related disciplines
Feature selection has a price in accuracy for efficiency
Researchers need to have the bigger picture in mind, not just doing selection for the purpose of feature selection.