Visual Analytics of Temporal Event Sequences

Temporal event sequence data (such as event logs) is collected in a wide variety of domains ranging from healthcare to cyber security, vehicle fault diagnosis, population living activities, and web clickstream records. Visual analytics aims to obtain a summary or overview of the data to allow knowledge discovery and support the improvement of the process being studied. Despite the great advances in visual analytics of event data, two main gaps were found in the literature. First, existing visualisations provide an overview of event sequences where its level-of-detail can be transformed by drilling down certain elements, but do not provide dynamic levels of detail simultaneously across sequences and longitudinally. Second, current overviews of event data focus on the visual encoding of sequential patterns but present limitations when representing temporal and multivariate attributes: the attributes are not encoded in the overview or if present, these are oversimplified (e.g. using average values).

This thesis tackles both gaps by proposing a technique to build a multilevel and multivariate overview of temporal event sequences. The overview is multilevel as its level of granularity can be transformed across sequences (vertical level-of-detail) or longitudinally (horizontal level-of-detail), using hierarchical aggregation and a novel cluster data representation Align-Score-Simplify. By default, the overview shows an optimal number of sequence clusters obtained through the average silhouette width metric – then users are able to explore alternative optimal sequence clusterings. The vertical level-of-detail of the overview changes along with the number of clusters, whilst the horizontal level-of-detail refers to the level of summarisation applied to each cluster representation. The overview is multivariate as it allows to visualise event types in the overview using an EventBox, a novel visual encoding that aggregates temporal and multivariate attributes for a set of event occurrences of the same type. The overview allows the identification of trends and outliers involving multivariate attributes within and across clusters.

The proposed technique has been implemented into a visualisation system called Sequence Cluster Explorer (Sequen-C) that allows detail-on-demand exploration through three coordinated views, and the inspection of data attributes at cluster, unique sequence, and individual sequence level. The technique is demonstrated through four case studies using three different types of real-world datasets in the healthcare domain: patient flow, hospital admissions and prescription history, and calls made to the emergency services. The case studies show how the technique can aid experts in exploring and defining a set of pathways that best summarise the dataset, while exploring data attributes for selected patterns. Moreover, Sequen-C was evaluated with 13 non-expert users. The results indicate that the system Sequen-C can allow novice users to quickly familiarise with the proposed visualisations and successfully obtain insights from the data according to the objective analytic tasks. Furthermore, the results of the System Usability Scale questionnaire indicate that Sequen-C has a good usability level.