Ontology-Driven Business Process Intelligence ◾ 415
Figure 15.13 Highlighted trace clusters. Four VRU output values are selected
as highlighted trace clusters. (a) Highlighted clusters on flow between services.
(b)Same highlighted clusters on flow between organizational units.
416 ◾ Jon Espen Ingvaldsen
processes. is group highlighting shows that VRU outputs consistently decide
which manual services will follow and which teams cover dierent areas of cus-
tomer service.
15.4.3.2 Evolution Charts
e presence of time stamp information in events and traces makes it possible to
animate the evolution of property values over time. Evolution charts animate how
numerical properties related to trace clusters develop over a period of time. A stan-
dard scatter plot shows data distributions across two dimensions (X-axis and Y-axis).
Evolution charts, on the other hand, reveal data across ve dimensions (X-axis, Y-axis,
color, size, and time). ey were inspired by the TrendAnalyzer from GapMinder
and use such animated visualizations to map world data (GapMinder 2010).
Based on concepts in the ontology and their internal ow relationships, EVS
constructs a large set of features that are available for statistics and data mining.
Duration measures are one such set of features that can be valuable for describing
trace cluster characteristics. e event log typically describes traces with dierent
proles. Traces cover dierent business areas; some overlap partly or completely
with each other, while others are separated. As a result, we must specify start and
end points when we describe duration values. When the event log is read into the
search index, the system remembers all possible starting points and end points for
successive event executions. ese combinations are used to construct and present
a set of possible duration features. An example of a duration feature is the time
elapsed from VRU complete to Regular Service start. Another set of features counts
the number of occurrences for each ontological concept. Two examples are the
number of times Martin or VRU complete was involved in the traces.
Each trace cluster is visualized as a bubble whose radius is a function of cluster
size (number of traces contained in cluster). e vertical and horizontal positions
of the trace cluster bubble depend on numerical features selected by the user, who
can also specify the animation speed, aggregation period, and number of animation
steps between the rst and last event in the visualized dataset.
Evolution charts make it possible to perform high level analyses such as identifying
long-term trends, and targeted analyses for gauging the impacts of a specic change.
An example of a high level analysis is investigating how business processes measures
(like durations and loads) develop relative to each other and over time in dierent
organizational units. Targeted analysis, on the other hand, may investigate how these
organizational units develop before and after a large marketing campaign.
Figure15.14 shows two screenshots from an animation sequence in an evolution
chart. We selected organizational teams as the highlighted clusters. Each cluster
is represented as a colored bubble. e discovered model in Figure15.11 revealed
that Regular Service is the most commonly requested service. In the chart, we set
the duration between Regular Service start and Regular Service complete to indi-
cate time spent servicing a customer as a horizontal feature, and duration between
Ontology-Driven Business Process Intelligence ◾ 417
VRU complete and Regular Service start to show response time before the phone is
answered by an operator as a vertical feature. We specied the animation to break
the time period into seven key frames (one per day), and calculate average values
based on the period between current animation time and 1 day back.
e two animation snapshots show how the horizontal and vertical feature values
change over time and move the bubble locations. e cluster diameters (dened by
the cluster sizes) have changed slightly over this period. On January 5, Team 5 had
a service time for regular services of 67 seconds, a response time of 29 seconds, and
processed a total of 24 customer calls. On January 7, their service time almost doubled
to 126 seconds and response time dropped below 4 seconds; they processed a total of
16 customer calls. We can see how Team 5 scored relative to the other teams. As the
animation screenshots show, a lot of volatility appears when we extract and visualize
customer service measures on a daily basis. With an event log containing data for lon-
ger time periods, the same analysis would be able to show longer trends and pinpoint
teams that are typically slow to answer the phone and start servicing customers.
15.4.3.3 Data Mining Characteristics
e data mining module has the potential of discovering unknown and valuable
patterns that characterize the highlighted trace clusters. Like evolution charts, data
Figure 15.14 Two screenshots of evolution chart module. (a) Snapshot of anima-
tion on January 5. (b) Snapshot of same data on January 7.
418 ◾ Jon Espen Ingvaldsen
mining modules select feature candidates from the ontological concepts and their
internal ow relationships. One feature dierent from evolution charts is that a
data mining module operates on both numerical and categorical features. One set
of common categorical features is based on ontological elements in the traces. In
EVS, such features are considered Boolean if they have a true (1) or false (0) value.
Figure15.15 shows how potential features are presented. A user can navigate
through this hierarchy and select those feature candidates that should be included in
Figure 15.15 Screenshot showing hierarchy of potential data mining features.
Ontology-Driven Business Process Intelligence ◾ 419
the data mining algorithms. In the gure, the user selected all duration features and
all service concepts as potential features to describe traces carried out by Team 1.
e ability to manually select a set of feature candidates gives a user control
over the solution space and the types of descriptions the data mining operation can
produce. is is useful to prevent the data mining algorithm from constructing
rules that are obvious, redundant, or of little knowledge value.
After a user nishes feature selection, a dataset is created. Figure15.16 shows a
fraction with the rst six traces of a bigger dataset. Each column represents a feature
and the feature name is written vertically in the column headers. e traces are rep-
resented as rows with a set of feature values. is dataset shows examples of duration
features and values. Column 2 represents a feature named duration, vru complete →
new customer start that describes the time elapsed between the rst occurrence of
VRU complete and the rst following occurrence of New Customer start. Only the rst
‘duration, vru start -> vru complete’
‘duration, vru complete -> new customer start’
‘duration, vru complete -> new customer complete’
‘duration, service in english start -> service in english complete’
‘duration, vru start -> stock trading complete’
‘team 2’
loren
‘it department’
michael
Eric
david
administration
sally
sharon
class
13000 1000 25000 ? ? 1 0 1 0 0 0 1 0 0 0
6000 ? ? ? ? … 1 0 1 0 0 0 1 0 0 0
13000 ? ? ? ? 0 0 1 0 0 0 1 0 0 0
5000 ? ? ? ? 0 0 1 0 0 1 1 0 0 0
5000 ? ? ? 183000 0 0 1 0 0 0 1 0 0 1
5000 ? ? ? ? 1 0 1 0 0 0 1 0 0 0
Figure 15.16 Fraction of dataset.
..................Content has been hidden....................
You can't read the all page of ebook, please click here login for view all page.