REPOSITORY > RESULTS

Abstract

This thesis presents an approach for the abstraction, exploration, interactive visualization
and interpretation of large multivariate time series. Interesting patterns in such datasets
often appear as periodic or recurrent behavior often caused by the interaction between the
variables.
To identify such patterns, we summarize the data as conceptual states, modeling temporal
dynamics as transition between the states. States are abstractions which summarize
all the measurements in a particular region of the ambient space. Transitions, with the
associated states, summarize the path made by the time series. This representation can
visualize large datasets with potentially billions of examples.
We extend the representation to multiple spatial granularities, which we call scales,
allowing the user to find patterns on several distinct scales. The spatial granularity of
the states can highlight certain patterns while obscuring others. Therefore, we design a
multi-scale representation, allowing patterns to emerge from within coarser scale states or
as the interaction between different parts of several states.
The result is an interactive web-based visualization tool called StreamStory. Stream-
Story couples the abstract representation with several tools that map the abstractions
back to domain-specific concepts using techniques from statistics, data mining and machine
learning. It is aimed at users who are not experts in data analytics, minimizing the
number of parameters to configure out-of-the-box.
We base the thesis on three main hypotheses: (i) combining techniques from machine
learning, data mining and statistics into a tool for data abstraction and visualization
can highlight the long-term behavior of a dataset and expose recurrent patterns while
mitigating cluttering effects, (ii) long-term and recurrent patterns can emerge at different
spatial granularities (i.e. scales) and traversing the scales interactively provides an effective
way to identify where such patterns emerge and (iii) the visual abstractions that summarize
the structure and dynamics of a dataset can be mapped back to the data and interpreted
as domain-specific concepts with quantitative data-driven visual tools.
We use four real-world datasets to confirm these hypotheses. With all the datasets
we provide an interpretation demonstrating how to map the abstractions back to domainspecific
concepts. We use weather and traffic datasets to demonstrate how to find and
interpret the long-term behavior of the data and a wind dataset to demonstrate how
traversing the scales can reveal the structure in the data. We also present three patterns
which are typical when exploring data using StreamStory with a GPS dataset. We find
and interpret several known as well as previously unknown patterns in these datasets.