Modelling Relational Data Mining

The goal of data mining is to discover new knowledge in the data. This thesis studies a number of relational data mining problems and demonstrates how they can be modelled and solved. Relational data mining involves dealing with complex and interconnected data, such as spreadsheets or relational tables in databases. Specifically, we follow a general, declarative, view on a class of relational data mining problems using the principles of declarative programming. In a declarative approach one specifies what the problem is, instead of how to solve it. The contributions of this thesis are: 1) it shows how Boolean Matrix Factorization can be generalized to the relational setting of Relational Data Factorization, and demonstrates how Relational Data Factorization can model a wide range of data mining problems; 2) it introduces the novel problem of Tabular Constraint Learning, where one recovers Excel-like formulae in spreadsheets; 3) it introduces the problem of Sketched Answer Set Programming, which allows one to mark parts of a logic program as uncertain, or open; then recover, or synthesize, the program using a number of positive and negative examples; 4) it demonstrates how our approach can be applied to model various structured and unstructured pattern mining problems under constraints. To summarize, we investigate how a number of important relational data mining problems can be modelled using Answer Set Programming and FO(.).

Publication year:2018

Accessibility:Open