Descriptive analysis of movement data based on multivariate movement patterns

Abstract

Descriptive analysis of data is an essential practice as it unravels the complexities of data into human knowledge. One type of data can be the object of descriptive analysis is movement data. Descriptive analysis of movement data when it involves multivariate movement patterns suffers from certain shortcomings. For example, in its current state, it is ineffective in the presence of many movement variables regardless of the data being labeled or unlabeled. On a separate note, although the interpretation of predictive models can produce patterns in form of interaction between variables, no technique is available in cases where data is unlabeled. Finally, model interpretation on movement data itself calls for improvement in three departments. First, the current model interpretation tools are ineffective in higher dimensions. Second, movement variables can be hard to interpret. Finally, predictive models are ineffective in revealing weak and shared (between classes) patterns. Resorting to a number of tools, in particular Feature Engineering, Outlier Detection, Topological Data Analysis, and Association Rule Mining, this thesis introduces a collection of novel pipelines to resolve the mentioned shortcomings. Various experiments with datasets of diverse nature demonstrate that the introduced pipelines are effective in resolving the target shortcomings. In particular, this thesis was able to identify the difference of behavior in ship classes; find multivariate interaction patterns among ships, cyclones, foxes, and soccer players; and find movement patterns in the mentioned datasets for the majority of data instances when the datasets are high-dimensional.