Multi-Relational Data Mining A Comprehensive Survey

2.0 Multi-Relational Data Mining

Multi relational data mining sub-domain has emerged to develop novel algorithms and techniques designed to work in relational database environments. During the last two decades, two different approaches have risen to address the challenging task of mining multi-relational structural data; propositionalization and upgrading.

Propositionalization is a process that leads from relational data and background knowledge to a single-table representation (Krogel, 2005). The literature of MRDM suggests two different techniques to achieve propositionalization; joining and aggregation. The joining technique is to use SQL join commands to link all related tables together in a single view, while aggregation technique is to summarize the raw data from multiple relations into a single compact table using SQL grouping commands. Both techniques will generate a universal table that is usually large and contain multiple nulls and redundant values. In addition to that, joining and aggregation techniques cause the loss of important semantic information that is usually represented in relational structure. Although propositionalization techniques are widely used in the domain of KDD, it also has significant drawbacks in terms of data representation and pre-processing cost.