Skip to main content

Data Mining

W. LiuD. Bell

Knowledge acquisition is expensive and often there is no expert around from whom to elicit the knowledge. We study Machine Learning and Data Mining techniques that allow models that are easily comprehensible to humans to be constructed from raw data.

Our earliest work in this area started with developing learning algorithms for constructing Bayesian Networks from data. Recent works in this area are strongly influenced by emerging real world applications, including:

  1. Developing anomaly detection algorithms for detecting abnormal behaviors in physical access control environment under the context of security within CSIT; developing graph-theory based algorithms for identifying exercise patterns and influences among participants in events; and developing social connection patterns from social networks.
  2. Developing various data analytical approaches, in collaboration with Belfast City Council, for analyzing data on Pollution, Waste disposal, Treatment, and Recycling; Anti-Social Behaviors, Buildings and Trees, etc.
  3. Developing real-time threats and anomaly prediction/detection algorithms, using knowledge discovered above, to provide real-time situation awareness for decision support.
Previous work within this area includes:
  1. We have carried out many projects on Rough sets - often to find rules which help in decision support. 
  2. Applications are in Nuclear Safety, Robotics, and Video/Image Mining. We also work on methods of generating and using Belief Networks, where a probabilistic or evidential graphical model to predict the behavior of systems.

Both theoretical and practical outputs have been produced. Eg. we have recently shown Rough Sets to be equivalent to, but more easily understood and used than, Association Rules. In the generation and use of Belief Networks, our contribution was to introduce the idea of monotone DAG faithfulness, which permits learning to happen in polynomial time when given a mutual-information oracle, and discussion of which has been taken up widely in the literature. We have carried out many applications including some in text mining, nuclear safety, telecommunications, and medicine. The methods were developed further and applied in the EU Framework ICONS and MISSION projects. We also applied our methods in robotics for which we won the BCS Machine Intelligence Prize in 2005.