Abstract

When something goes wrong in an organisation, analysts and domain experts work hand in hand to find root causes, led by their experience and beliefs. In other words, they typically only find what they are looking for. Successful root cause analyses therefore require creativity, time and a portion of luck.

Modern data mining techniques address this problem by exploring large search spaces in a short time, producing a set of potentially interesting patterns. However, machines cannot judge how interesting a pattern is. Analysts may find the results disappointing, because they still have to go through a large number of patterns, some of which they will know and others they are not interested in.

The presentation will introduce our own techniques for pattern mining over time and address the following questions:

1. How can a machine judge how interesting a pattern is?
2. How do we include time as a dimension?
3. How can we present, visualise and explore data mining results?
4. What would a truly interactive data mining system look like that combines the speed of machine computation with the experience and domain expertise of analysts?