Classifiers

1
Classifiers

• Classification is a process of two steps
• _ Learning training data are analyzed by a
  classification algorithm. Build a set of rules to
  describe the set of data.
• _ Classification given an instance, rules are
  used to classify the instance under one of
  classes.
• Rules Pattern -gt class label
• Pattern used
• Emerging patterns
• Generators
• Closed patterns
• Non-redundant discriminative patterns

2
Classifiers
Given a test instance, a top K rules from each
class are matched to the test instance. A score
is computed for class i Where
is the support of X in Di. The test instance will
be classified under class whose score is bigger.
3
Architecture
Patterns miner
Produce frequent patterns
Results summary
Classifiers

• Analysis program

Produce classifications
Graphs
Input
Output file

• Experiment program
• Read the dataset
• Devide into training and testing data
• Send input to classifier
• Get results and compute the accuracy.

output
USER
4
Architecture

• Input pattern miners (DPM, oddratio, ..) and
  format converters.
• Run experiment
• Parameters are specified in a text file
• The program will take all combination of
  parameters to run
• Run analysis program
• Four variables to be analyzed
  threshold, missing data and delta and type of
  patterns.
• View summary and graph
• Graph is plot using gnuplot

5
Experiments

• The datasets are from UCI repository
• diabetes
• glass
• iris
• mushroom
• vote
• zoo
• Continuous data were stratified 10-fold except
  glass 20-fold.
• Missing value are set at 0, 5, 10, 20, 30 ,50 .

6
Results
7
Remark

• Generators are slightly better than closed
  patterns.
• In compare with PCL, we are more accurate at
  glass but less accurate at zoo and vote.
• Missing data has less effect on the accuracy.