Database Reference
In-Depth Information
tables, which, once queried, will trigger the execution of mining algorithms. This is
an example:
Here,
Sets
,
Rules
and
Concepts
are virtual mining views.
A limitation of most SQL-based approaches is however that they are limited
to itemset patterns or association rules. How to specify graph mining or sequence
mining tasks in this context is still an open question. Most constraint-based graph
mining or sequence mining systems currently use special purpose languages.
The general idea of linking constraint-based mining to database querying has been
studied in the area of
inductive databases
and
inductive querying
[
9
,
16
].
Constraint Programming
Constraint-based mining has many similarities to
generic constraint satisfaction problem (CSP) solving as studied in the Artificial In-
telligence (AI) community. Both areas essentially require the discovery of solutions
in a space of possible solutions satisfying constraints. To deal with generic CSPs,
the AI community has developed generic systems known as
constraint program-
ming systems
. These systems provide languages in which programmers can specify
constraint satisfaction problems; statements in these languages can be solved by
various types of solvers, including generic propagation-based solvers. As we have
seen earlier, many depth-first constraint-based itemset mining systems are also based
on propagation, and hence it is not surprising that generic constraint-based itemset
mining fits naturally into a constraint programming context as well.
This observation was used by Guns et al. to formalize constraint-based item-
set mining tasks in generic constraint programming languages [
13
,
14
]. This is
an example in the most recent version of the MiniZinc constraint programming
language:
It specifies the task of frequent itemset mining;
cover
is a function available in a
MiniZinc library, implemented in the MiniZinc language ifself as well.
Statements in the MiniZinc language can be executed by a generic constraint
programming system, or by a specialized data mining system, if one exists [
14
].
However, it was shown that generic constraint programming systems implement
many types of propagation automatically, and hence that specialized systems are
often not needed if a task can be modelled in the MiniZinc language.
Similar to the SQL-based languages, it is at this moment not understood how
to integrate graph mining or sequence mining tasks in an elegant matter in the CP
setting.
