Adaptive Methods for Spatiotemporal Stream Data Mining
| aut.embargo | No | en_NZ |
| aut.thirdpc.contains | No | en_NZ |
| dc.contributor.advisor | Pears, Russel | |
| dc.contributor.advisor | Kasabov, Nikola | |
| dc.contributor.author | Hartono, Reggio Nurtanio | |
| dc.date.accessioned | 2018-11-20T23:05:15Z | |
| dc.date.available | 2018-11-20T23:05:15Z | |
| dc.date.copyright | 2018 | |
| dc.date.issued | 2018 | |
| dc.date.updated | 2018-11-20T13:45:35Z | |
| dc.description.abstract | The availability of temporal and spatiotemporal data is increasing, and the use of traditional statistical techniques to deal with such data is insufficient. Novel methods that are capable of adapting to changing patterns in time-variant spatiotemporal data need to be developed. To achieve this objective, the thesis proposes three different methods to deal with various types of data and employ distinct approaches to tackle common problems faced in spatiotemporal data mining. The first method deals with multiple time series and presents a development of a generic framework to extract knowledge in the form of temporal rules. The main component is a modified association rule mining algorithm that also works with time dimension, producing rules in the form of A ⇒ B. As part of the research, a discretisation technique inspired by concept drift detection is also proposed. The framework was then applied to a dataset that tracks the number of aphids caught in traps along with weather variables over almost twenty years in the Lincoln region in Canterbury, New Zealand. The second method deals with building local models for a time-step ahead spatial prediction problem. Taking advantage of the locality-preserving property of the space-filling Hilbert curve, the method is able to work with existing concept drift detection algorithms to automatically determine where and when in the spatiotemporal landscape that patterns are changing. The framework was tested on the earthquake catalogue data around the Christchurch region. The empirical results reveal that the local models improved the prediction accuracy of up to 9% on one of the tests when compared to a standard incremental model building approach based on a fixed size sliding window scheme. The third method employs a Spiking Neural Network (SNN)-based system called NeuCube to build an early event prediction system. The system was trained to differentiate the seismicity readings obtained from spatially scattered seismograms around Canterbury both before large earthquakes happen and periods of low seismicity between 2010 and 2016. The system was tested to examine whether NeuCube could learn from complex data and demonstrate a capability of predicting large earthquakes with a reasonable window of time. The results from this scheme are promising as NeuCube could predict major seismic events with a much higher true positive rate (0.91) while keeping the false positive rate significantly lower (0.08) when compared to other prediction algorithms. | en_NZ |
| dc.identifier.uri | https://hdl.handle.net/10292/12054 | |
| dc.language.iso | en | en_NZ |
| dc.publisher | Auckland University of Technology | |
| dc.rights.accessrights | OpenAccess | |
| dc.subject | Spatiotemporal | en_NZ |
| dc.subject | Data mining | en_NZ |
| dc.subject | Spiking neural networks | en_NZ |
| dc.subject | Earthquakes | en_NZ |
| dc.title | Adaptive Methods for Spatiotemporal Stream Data Mining | en_NZ |
| dc.type | Thesis | en_NZ |
| thesis.degree.grantor | Auckland University of Technology | |
| thesis.degree.level | Doctoral Theses | |
| thesis.degree.name | Doctor of Philosophy | en_NZ |