Detecting and describing non-trivial outliers using Bayesian networks

Traditionally, outlier detection is the task of discovering highly deviated objects. However, mere discovery of outliers may not be sufficient for an application to be successful. Verification on genuineness of the reported outlier, and understanding on its exceptional properties are important to be integrated in the discovery process. This research proposes an approach to differentiate among non-trivial, strong, weak and trivial outliers using domain knowledge captured by a Bayesian network. The approach also provides an environment to explain and describe non-trivial and strong outliers using Bayesian framework. Bayesian networks are very useful in computing probability of an event. In this work, those observations are identified which are less likely to fit into the relationship that exist between variables encoded in the graphical structure of the model. Encouraging preliminary experimental results supports use of Bayesian approach for outlier detection and description in diverse application areas.