A Closer Look at WEKA

1
A Closer Look at WEKA

• Notes on Manipulating Data in Weka

2
General Remarks

• WEKA is a java based system which allows
  contribution by independent researchers.
• Although commercial use is not prohibited, it is
  primarily a research oriented tool, which results
  in the difficulties encountered in the interface
  (input, output).

3
Preprocessing (Input)

• WEKA uses arff files for input. It is a
  text-based format. See the example files under
  installation folders.
• It accepts comma delimited text files as well.
  Given an Excel file, this formatting requirement
  can be easily met Save as (csv, comma delimited
  for file type)

4
Preprocessing (Input)

• If one wishes to modify the input file, for
  instance remove some attributes, there are two
  ways to proceed
• Modify the csv file using Excel
• Modify the input file using WEKA filters and save
  the file as an arff file for future use.
• WEKA filters are strong and capable tools, and
  their use is documented in the users manual.

5
Applying a Filter

• Select a filter, click on add. When apply
  filters button is pressed, all filters that
  appear highlighted in the box will be applied.
• Once a file has been filtered, it is technically
  a different file (although it might be the same
  as the original in terms of content) and can be
  saved as such.
• The saved file needs to be reloaded for the
  filtered data to be in use.

6
Some Basic Filters

• Add filter adds an attribute to dataset
• Attribute filter removes attributes from the
  data set
• Attribute expression creates new numeric
  attribute as a function of an existing one
• Discretize filter converts numeric attribute
  into nominal by using bins
• Randomize filter randomizes order of instances

7
Conclusion

• The GUI of WEKA has improved a lot.
• From an instructional perspective, it is a strong
  and useful tool because it supports alternative
  methods.
• For more information
• http//www.cs.waikato.ac.nz/ml/weka/

8
Some Classifiers in WEKA

• zeroR predicts the majority class, or the
  average if the output attribute is numeric
• oneR uses one attribute only in rule generation
• NaiveBayes implements the Bayes classifier (the
  normal distribution assumption for numerical
  attributes can be modified if necessary)

9
Some Classifiers in WEKA- ctd

• IBk instance based k-nearest neighbor classifier
• J4.8 (J4.8 PART) Decision tree-based algorithm,
  PART generates rules
• ID3
• Logistic Regression
• ...

10
Methods for Prediction (Numerical Output
Attribute)

• Regression
• Regression / Model trees (M5)
• Neural Networks (supports classification as well)

11
Other Methods

• Clustering
• K-means
• EM
• Cobweb
• Association (A Priori Algorithm)