Discretisation of conditions in decision rules induced for continuous data.

Area of Science:

• Computer Science
• Data Mining
• Machine Learning

Background:

• Traditional data discretisation occurs before data mining, potentially discarding valuable information.
• Existing methods often lead to conclusions based on incomplete or reduced datasets.

Purpose of the Study:

• To propose and evaluate a novel approach of performing data discretisation after knowledge mining.
• To investigate the impact of post-hoc discretisation on the size and quality of rule classifiers.
• To assess the computational feasibility of this method in stylometric analysis.

Main Methods:

• Inducing decision rules from real-valued features.
• Discretising datasets using categories derived from inferred rules.
• Translating rule conditions into a discrete domain.
• Evaluating rule classifiers in stylometric analysis of text.

Main Results:

• The proposed method significantly reduces the size of rule sets.
• Predictive accuracy of classifiers is maintained despite size reduction.
• Multiple data discretisation methods can be tested efficiently.

Conclusions:

• Post-hoc data discretisation is an effective strategy for enhancing data mining efficiency.
• This approach offers a balance between rule set size reduction and predictive performance.
• The method is computationally viable for complex applications like stylometric analysis.