Michael Berthold: Abstract

The focus for data analysis has shifted quite a bit from its traditional origins in recent years. In many applications, the generation of data has never been a problem but the real problem of generating reasonably well annotated data has increasingly become the true bottleneck. However, in order to apply traditional data analysis methods a solid understanding of the patterns to be extracted is required.

Increasingly, modern setups require to take back one step and use explorative systems that allow to create first insights into vast and heterogenuous information repositories which will then actually lead to more concrete questions that can be addressed by more traditional data analysis techniques. Therefore instead of providing a general overview (which often does not reveal any interesting insights) systems need to be developed that can quickly focus the user\'s attention on potentially interesting, more local artefacts which can then be analysed (or interactively explored) \"sideways\", accompanied by zoom-in/zoom-out actions but also in concert with the already existing extensive reservoir of data analysis techniques.

In this talk we will describe a framework using hetergenuous information networks to fuse information from various, diverse data sources. The data can be semantically meaningful (such as ontologies) but it can also be simply evidential (such as experimental data revealing mere co-occurence statistics or other measures of correlation). Formal network analysis does often not allow to discover sufficiently complex sub networks. We will therefore describe a number of more scalble, heuristic methods to find nodes, paths, and subnetworks of (potential) interest, revealing potential areas of synergy or previously unknown links indicating potential concepts that are of interest or entirely missing.

Organised by BCS SGAI The Specialist Group on Artificial Intelligence http://www.bcs-sgai.org