Motivation for the RILA project was to develop a data mining software that can apply supervised machine learning algorithms to multiple connected tables in relational databases, without requiring the input tables joined.
In a simple usage scenario;- We have list of tables to be analysed that are connected to each other through foreign key relations
- One of the central tables specified as hub-table should have one or more primary keys to uniquely identify objects being analysed
- One table selected as class table and one of its columns selected as class attribute
Patterns found by RILA can either be used in predicting unseen objects attribute values (by comparing the attribute values recorded by the patterns/rules), or to explore the relationships between object attribute values. The second way of using relational rules can also be described as data summarisation.
Basic details of RILA was first described in a conference paper in 2003 [1]. Later a new rule selection strategy, called select late, was developed [2]. Rila was tested using PostgreSQL, Oracle and MySQL databases.
Although the project was started more than 10 years ago now, it is not in a very good shape and still missing a downloadable distribution.
Publications and documents:
- A multi-relational rule discovery system, Lecture Notes in Computer Science, Volume 2869/2003
- A new relational learning system using novel rule selection strategies, Knowledge-Based Systems Volume 19, Issue 8, December 2006