Good and bad practices in propositionalisation

Data is mainly available in relational formats, so relational data mining receives a lot of interest. Propositionalisation consists in changing the representation of relational data in order to apply usual attribute-value learning systems. Data mining practitioners are not necessarily aware of existing works and try to propositionalise by hand. Unfortunately there exists some tempting pitfalls. This article aims at bridging the gap between data mining practitioners and relational data, pointing out the most usual traps and proposing correct approaches to propositionalisation. Similar situations with sequential data and the multiple-instance problem are also covered. Finally the strengths and weaknesses of propositionalisation are listed.