[Next] [Previous] [Top]

Tutorial/Practical

Example

Clementine

• Supplied by Integral Solutions Limited (ISL), Basingstoke, England

• Visual Programming Interface
  • builds a discovery model

  • performs learning task

• Uses neural networks and rule induction

• Data sources
  • ASCII file format, Oracle, Informix, Sybase and Ingres

• Clementine has many useful facilities:
  • Data Manipulation - construct new data items derived from existing ones, and breaking the data down into meaningful sub-sets

  • Browsing and Visualisation - displaying aspects of the data using interactive graphics

  • Statistics - confirming suspected relationships between factors in the data

  • Hypothesis testing - constructing models of how the data behaves and verifying them

Clementine Example

Drug trials

• A number of hospital patients all suffering from the same illness were treated with a range of drugs

• 5 different drugs were available and the different patients responded differently to the different drugs

• Problem - which drug is appropriate for any given future patient?

Problem solving

Stages

• Accessing the data
  • read in the data e.g. from a text file with delimiters

  • name the fields
    • age

    • sex

    • BP - High, Normal or Low

    • Cholesterol - Normal or High

    • Na - blood sodium concentration

    • K - blood potassium concentration

    • drug - i.e. to which the patient responded

• View the records by using the table node e.g.

• Can select fields or filter the data

• Display properties of the data e.g.
  • what proportion of cases respond to each drug?

• Answer DrugY followed by DrugX

• Finding a relationship e.g.
  • relationship between sodium and potassium levels as displayed in a point plot
    
    

  • Random scattering - no apparent relationship

• Re-examine according to a particular drug i.e. drugY, and sodium to potassium ratio

• calculate the Na/K ratio i.e. as a derived field or node
  
  

• Patients with a high Na to K ratio respond best to drugY

Machine learning

Clementine

• Which is the best drug for any patient?
  • Filter the unwanted fields

  • Define types for the fields

  • Building rules and training nets i.e. by attaching the appropriate nodes

Building rules and training nets

• Net trained, rules built on 200 example cases

Net and rules completed

Rules formulated

• The rules first decision is based on the same criterion discovered previously i.e. allocates drugY to patients with a high Na to K ratio

UUJ Example

House price prediction

• Problem - mass appraisal of property in N.Ireland

• 10 attributes per property including:
  • Ward No, Area No, Price, Age of house, Number of bedrooms, Detached/Semi, Type of building - (bungalow, house, chalet etc.), Heating

• Best predictive accuracy
  • 82% with mean error 15%

• After removal of outliers
  • 93% with mean error 7.8%

Data file - houses.dat

Graphical Output

• Outliers - any property over 100,000 had land attached

Statistics Produced

Predictions set

Clementine stream

Clementine Output

Rule Browser

REFERENCES

• Knowledge Discovery Data Mine
  • http://info.gte.com/~kdd/

• University of Ulster Jordanstown
  • Database Mining Interest Group (UUDMIG)

  • http://iserve1.infj.ulst.ac.uk/main.html

• Queens University Parallel Computer Centre
  • Training and Education - course materials

  • http://www.pcc.qub.ac.uk/

• Articles
  • Byte, October 1995