Abstract

Data clustering can be used as a tool for data mining, knowledge discovery and intelligent data analysis. This thesis reviews density-based, partitional and hierarchical clustering methods for classifying data clusters. Outliers are often classed as noise, but in Flight Data Analysis they can be used to highlight flights during which incidents are more likely to occur. Reliable assumptions of the underlying data format cannot be made a priori without risk of mis-identifying valid abnormal data which does not fit the expected data format - this may include outliers. Machine Learning techniques are used over more tra-ditional Statistical approaches. After experimentation of crisp (exclu-sive membership to a cluster) and fuzzy (shared membership to many clusters) data clustering methods upon flight parameters during the descent, the Fuzzy clustering by Local Approximation of MEmberships (FLAME) (Fu & Medico, 2007) clustering method provided the best fit to the dataset as it was best able to separate objects belonging to clusters and those belonging to outlier classes. Various cluster valid-ity indexes have been reviewed and tested upon different clustering algorithms.
This thesis proposes a novel scaling algorithm to aggregate three clus-ter validity indexes which determines the most appropriate a priori input parameter values for the FLAME clustering algorithm for a given dataset. The method exaggerates the strengths of each validity index, proportioning the aggregate validity result according to best fit. This thesis discusses the results and recommends future work includ-ing the application multi-dimensional visualisation techniques which show promise in improving understanding of contributing factors of outlier objects.

Additional Notes

Supervisors: Dr David Brown and Eric Grummitt

Course: Master of Philosophy - MPhil

Date Deposited: 2017-05-10

URI/permalink: https://library.port.ac.uk/dissert/dis12876.html