Mining Relational Association Rules for Propositional Classification

In traditional classification setting, training data are represented as a single table, where each row corresponds to an example and each column to a predictor variable or the target variable. However, this prepositional (feature-based) representation is quite restrictive when data are organized into several tables of a database. In principle, relational data can be transformed into prepositional one by constructing prepositional features and performing classification according to some robust and well-known prepositional classification methods. Since prepositional features should capture relational properties of examples, multi-relational association rules can be adopted in feature construction. Propositionalisation based on relational association rules discovery is implemented in a relational classification framework, named MSRC, tightly integrated with a relational database. It performs the classification at different granularity levels and takes advantage from domain specific knowledge in form of hierarchies and rules. In addition, a feature reduction algorithm is integrated to remove redundant features. An application in classification of real-world geo-referenced census data analysis is reported.