Discovery of Driving Behavior Patterns - Smart Information Systems: Computational Intelligence for Real-Life Applications - page 311

and emission evaluations. Due to the fact that Steven's approach is less expensive
and time consuming than a nationwide survey, the decision committee felt positive
about his idea and assigned him to lead the research project.
Leading a team of researchers in developing more efficient and environment-
friendly combustion engines does not only mean a real breakthrough in Steven's
career, but also a huge success on a personal level. Since his childhood, he always
dreamed of finding a way to use own skills to do something for the benefit of the
nature his father taught him to love. This is a unique opportunity for Steven to make
a positive impact on the environment of future generations. He wants his children
and grandchildren to enjoy and experience nature in the same was as he did as a kid.
With this in mind, Steven accepts the challenge of his lifetime.
12.1 Introduction
Clustering of times series data is of pivotal importance in various applications [ 9 ]
such as, for example, seasonality patterns in retail [ 13 ], electricity usage profiles [ 17 ],
DNA microarrays [ 26 ], and fMRI brain activity mappings [ 39 ]. A crucial design
decision of any clustering algorithm is the choice of (dis)similarity function [ 1 , 14 ].
In many clustering applications, the underlying (dis)similarity function measures
the cost of aligning time series to one another. Typical examples of such functions
include the DTW and the Euclidean distance [ 4 , 10 , 27 ].
Alignment-based (dis)similarity functions, however, seem not to be justified for
applications, where two time series are considered to be similar, if they share common
or similar subsequences of variable length at arbitrary positions [ 2 , 16 , 28 , 40 ].
A real-life example for such an application comes from the automotive industry,
where test drives of vehicles are considered to be similar, if they share similar driving
behavior patterns, i.e., engine behavior or drive maneuvers, which are described by
the progression of multiple vehicle parameters over a certain period of time [ 33 , 35 ].
In this scenario, the order of the driving behavior patterns does not matter [ 32 ], but
the frequency with which the patterns occur in the contrasted time series.
Recent work [ 5 ] on time series distancemeasures suggests to neglect irrelevant and
redundant time series segments, and to retrieve subsequences that best characterize
the real-life data. Although subsequence clustering is a tricky endeavor [ 12 ], several
studies [ 2 , 7 , 16 , 28 , 40 ] have demonstrated that in certain circumstances ignoring
sections of extraneous data and keeping intervals with high discriminative power
contributes to cluster centers that preserve the characteristics of the data sequences.
Related concepts that have been shown to improve clustering results include time
series motifs [ 2 , 16 ], shapelets [ 28 , 40 ], and discords [ 7 ].
In this contribution, we propose to adopt recurrence plots (RPs) [ 18 , 21 , 22 ] and
related recurrence quantification analysis (RQA) [ 19 , 20 , 23 ] to measure the simi-
larity between multivariate time series that contain segments of similar trajectories
at arbitrary positions in time [ 32 ]. We introduce the concept of joint cross recur-
rence plots (JCRPs), an extension of traditional RPs, to visualize and investigate