REPOSITORY > RESULTS

Abstract

Most data mining approaches focus on analysing data represented in tables with a fixed number of columns. More complex data storage formats include relational databases and information networks, for which standard data mining algorithms are not well suited. This thesis presents several contributions to the field of complex data mining.
The first part of the thesis addresses the problem of node classification in heterogeneous networks, i.e., networks containing different types of nodes and edges. We focus on a network decomposition algorithm that constructs one or more homogeneous networks from a heterogeneous network. We propose several heuristics for evaluating the significance of so-called intermediary nodes in heterogeneous networks. The motivation for these heuristics originated from methods used to construct bag-of-words vectors in text mining. The heuristics were designed to penalise the influence of intermediary nodes that connect too many nodes, as well as those that connect many nodes with different class labels. We tested the new heuristics on several datasets and demonstrated that their use improves the classification of network nodes. The proposed approach has been implemented in the newly developed HinMine methodology.
Second, the thesis presents our work on combining network analysis methods with semantic data mining (SDM) algorithms to explain datasets annotated with ontology terms. The proposed methodology, called NetSDM, uses network node-ranking methods to evaluate the significance of terms in the background knowledge ontology. We tested the methodology on several datasets, including a biological dataset containing gene expression data from potato plants infected with a virus. Our results show that NetSDM can accelerate the state-of-the-art semantic data mining algorithm Hedwig by a factor of 100, enabling it to discover high-quality rules describing a dataset in minutes rather than hours.
Third, we present a methodology for data visualisation using banded matrices. This methodology improves the interpretability of the outputs of semantic data mining algorithms by using banded matrix algorithms to uncover hidden structures in the data. Because these structures are discovered independently of the SDM algorithm, the results can be used to create an informative, easy-to-understand visual overlay of both sets of results.
Finally, the algorithms presented in the thesis have been made publicly available. The HinMine and NetSDM methodologies are available as workflows on the ClowdFlows online data mining platform, while the banded matrix algorithms are available as Python packages on GitHub.