The iZi Project: Easy Prototyping of Interesting Pattern Mining Algorithms
Résumé
In the last decade, many data mining tools have been developed.
They address most of the classical data mining problems such as
classification, clustering or pattern mining. However, providing classical
solutions for classical problems is not always sufficient.
This is especially true for pattern mining problems known to be “representable
as set”, an important class of problems which have many applications
such as in data mining, in databases, in artificial intelligence, or
in software engineering. A common idea is to say that solutions devised
so far for classical pattern mining problems, such as frequent itemset
mining, should be useful to answer these tasks. Unfortunately, it seems
rather optimistic to envision the application of most of publicly available
tools even for closely related problems.
In this context, the main contribution of this paper is to propose a modular
and efficient tool in which users can easily adapt and control several
pattern mining algorithms. From a theoretical point of view, this work
takes advantage of the common theoretical background of pattern mining
problems isomorphic to boolean lattices. This tool, a C++ library called
iZi, has been devised and applied to several problems such as itemset
mining, constraint mining in relational databases, and query rewriting
in data integration systems. According to our first results, the programs
obtained using the library have very interesting performance characteristics
regarding simplicity of their development. The library is open source
and freely available on the Web.