Scalable Decision Tree SPRINT

1
Scalable Decision Tree SPRINT

• Project Members
• Kaushal Mittal
• Abhishek Seth
• Amar Agrawal

2
Problem Statement

• Current decision tree implementation in Weka
  fails for large datasets.
• Scalable implementation of decision trees in
  Weka.
• Support for disk resident data.

3
Challenges

• Instance class in Weka loads the entire training
  data in memory.
• Multiple copies of the instance data made at
  several points during the training.
• Other classes assume the existence of
  memory-resident instance data.

4
Changes in Weka

• Extended the Instance class to support disk
  resident data.
• Use of cache and random access files.
• Changes to the Evaluation class to work with the
  new SInstance class.

5
Decision Tree Classifier

• Design similar to Weka classifier J48.
• SPRINT algorithm implemented.
• Use of disk resident attribute lists.
• Generates a binary classifier tree.
• Uses Gini index as split criteria.

6
Results

• Accuracy comparable to J48.
• Glass 214
• J48 - 100
• Sprint 91.667
• Adult
• J48 83.3
• Sprint 79.8
• Execution time More than default J48 for small
  data sets(IO). For large data sets, Weka fails.