The Self-Organising Map Network: An Interactive Visual Data Mining Framework for Exploratory Causal Analysis

Abstract

This thesis introduces a novel framework for analysing causality from complex data in diverse application domains, such as ecological and health areas. The framework, called SOM Network (SOMNet), allows complex domain data to be better analysed by empowering domain experts to generate analytical processes, explore associational patterns, and interpret causal information through an interactive visual data mining process. In the SOMNet, a weighting scheme based on the data similarity associates different datasets using different SOMs to define the indeterminate estimation of causality. Since analysing changes to weighted SOM patterns is essential for causal analysis, a visual analysis approach is developed by extracting change features and developing visualisation methods to facilitate global and local change comparisons. The proposed SOMNet is applied to a complex case study of discovering evidence-informed causal knowledge in a mental health domain. The SOMNet's capability is demonstrated by analysing the mental health domain data for health policy planning in a real-world context. The domain experts identify outliers in the pre-processing, while identifying and explaining global and local data patterns in the mid-processing of the data. Based on the pattern information, the indicator relationships between input and output datasets are examined and interpreted. The integrated information is used to guide further change analysis exploring an evidence-informed causal structure. Finally, the SOMNet is evaluated by the domain experts in the post-processing of the data. The study results show that the SOMNet approach is useful, relevant, effective and efficient in understanding highly complex domain data and discovering causal knowledge, and offers significant potential for decision support systems. Using the unique modelling of the SOMNet, it is possible to both dynamically communicate analytical information and discover potential causalities from complex domain data.