Learning If-Then Rules with I2D

A system for learning if-then rules from a dataset is available in the Feature Vector Editor. Perhaps the most successful AI algorithms are a family of entropy-based techniques for inducing decisions trees. Decision trees organize tests hierarchically such that running the tests along a path from the top to any leaf classifies an example within a category. Because they rely on global information measures, these techniques require a one-to-one mapping between values of compared variables that is found only in flat datamodels (one feature vector). Complex datamodels -- for example the hierarchically-structured feature vectors in the SHERFACS International Conflict Management Dataset -- almost always give rise to one-to-many mappings between values of compared variables, but this violates the fundamental assumption of conventional decision-tree learners, rendering them unusable for all but the most limited analyses of SHERFACS. Unseld and Mallery (1991) found a unifying solution to this problem that generalizes earlier techniques along several dimensions in order to handle complex datamodels. The key insight is to decompose examples into local observations until one-to-one comparisons again become possible, and then, to recombine these local comparisons to ultimately arrive at global information measures. Their Induction Interaction Detector (I2D) implements an inductive rule learner suitable for studying complex datasets like SHERFACS because it supports learning on the tree-structured temporally related feature-vector data.

Significant features include:

• Time-Dependent Analysis: Support for temporally structured data makes it possible to learn decision trees and rules to predict outcomes on the basis of prior variable values;

• Regularity Recognition: Instead of narrowly focusing on classification, I2D is designed to identify regularities between dependent and independent variables;

• Exploratory Data Analysis: Because it can identify regularities in data and is not limited by degrees of freedom, I2D is an excellent tool for exploratory data analysis, sifting through new or little-understood data to find the significant relationships that statistical methods might later test.