Multi-relational Data Mining: a perspective

Multi-relational data mining (MRDM) is a form of data mining operating on data stored in multiple database tables. While machine learning and data mining are traditionally concerned with learning from single tables, MRDM is required in domains where the data are highly structured. One approach to MRDM is to use a predicate-logical language like clausal logic or Prolog to represent and reason about structured objects, an approach which came to be known as inductive logic programming (ILP) [18, 19, 15, 16, 13, 17, 2, 5]. In this talk I will review recent developments that have led from ILP to the broader field of MRDM. Briefly, these developments include the following: - the use of other declarative languages, including functional and higher-order languages, to represent data and learned knowledge [9, 6, 1]; - a better understanding of knowledge representation issues, and the importance of data modelling in MRDM tasks [7, 11]; - a better understanding of the relation between MRDM and standard single-table learning, and how to upgrade single-table methods to MRDM or downgrade MRDM tasks to single-table ones (propositionalisation) [3, 12, 10, 14]; - the study of non-classificatory learning tasks, such as subgroup discovery and multi-relational association rule mining [8, 4, 21]; - the incorporation of ROC analysis and cost-sensitive classification [20].