Temporal Exceptional Model Mining Using Dynamic Bayesian Networks

The discovery of subsets of data that are characterized by models that differ significantly from the entire dataset, is the goal of exceptional model mining. With the increasing availability of temporal data, this task has clear relevance in discovering deviating temporal subprocesses that can bring insight into industrial processes, medical treatments, etc. As temporal data is often noisy, high-dimensional and has complex statistical dependencies, discovering such temporal subprocesses is challenging for current exceptional model mining methods. In this paper, we introduce Temporal Exceptional Model Mining to capture multiple and complex relationships among temporal variables of a dataset in a principled way. Our contributions are as follows: (i) we define the new task of temporal exceptional model mining; (ii) we characterize the discovery of exceptional temporal submodels using dynamic Bayesian networks by means of a new distance measure, (iii) we introduce a search procedure for exceptional dynamic Bayesian networks optimized by properties of the proposed distance, and (iv) the practical value of the proposed method is demonstrated based on simulated data and process data of funding applications and by comparisons with other exceptional model mining methods.