DatGen: To Do and Questions

Yet To Be Done

1. Write-out and read-in parameters into/from configuration files
   • Allows for repeatable runs and historical records
     
2. Write-out and read-in a rule-base
   • Allows for repeatable experiments
     
3. Support mapping column domains to files
   • Allows for custom domain values. Especially nominal domains such as colors and conditions.
     
4. Support for Background Knowledge
   • Attach generalization hierarchies to attributes of any data type. For continuous a tree decomposition could look like this (imagine the lines connecting the tokens):

                        ANY
   
           Low          Mid          High
   
       A        B        C        D        E
   
    [0-.9) [.9-1.7) [1.7-2.6) [2.6-4.1) [4.1,6]
   
   

5. Create rules which map to balanced decision tree
   • i.e. Murthy, Salzberg 95
   • changing the know. representation may be difficult
   • better to find a mapping from tree to DNF rules
     
6. Generate rules with more than one attribute class
   • e.g. Association rules (Agrawal et. al)
   • e.g. Bayesian networks
    
7. Support every model representation formalism used by the more important data mining algorithms.
   • Decision Lists
   • Centroids
   • Neural Networks
   • First Order Logic

Questions

1. How to find out what people need?
   
2. How will data mining benchmarks be defined?
    
   • Dynamic Definition: A general set of DatGen parameters is specified.
   • Fixed Rule Definition: A (set) of domain descriptions are proposed.
   • Success Factor: Whatever benchmark style is chosen it has to be tested against several learn-from-example algorithms and the results must expose strengths and weaknesses in each.