The Automatic Explanation of Multivariate Time Series

Due to the advances in data capture and storage techniques over the last decade, the size of Multivariate Time Series (MTS) data being recorded has grown massively. Many of these MTS are characterised by a large number of interdependent variables with large possible time lags. If new and useful knowledge is to be automatically learnt from this type of data in order to aid the understanding of the underlying processes, a paradigm must be identified that is capable of modelling data with these characteristics but at the same time exhibiting transparency in how it models the data. A key challenge is that the number of possible models is very large since it does not only depend on the number of time series variables, but also on the size of possible time lags between ‘causes’ and ‘effects’. In this thesis a general framework is described for automatically learning probabilistic models from MTS with large time lags and high dimensionality in order to explain the underlying processes involved. Specifically, a novel method to learn dynamic Bayesian networks for explanation from these series is developed. This involves an efficient pre-processing stage, which effectively groups MTS variables in order to reduce the dimensionality of the problem. After pre-processing, a combination of Evolutionary Programming, Genetic Algorithms and heuristics is used to speed up convergence when learning models. In addition, an approach is looked at for the off-line learning of dynamic Bayesian networks with changing dependency structures. All experiments have been carried out on a mixture of synthetic and real data taken from an oil refinery repository. The resultant models are used to generate explanations that are evaluated in several ways, including reviewing the feedback from chemical process engineers. These results have demonstrated that the proposed framework is very promising in terms of both efficiency and accuracy. Allan Tucker The Automatic Explanation of Multivariate Time Series Birkbeck College