Information Technology Reference
In-Depth Information
Fig. 3.
Example of unfiltered ACG (left side) and filtered ACG (right side)
as two correlated messages cannot be associated with the same node in the ACG
and therefore, if a large number of messages are correlated with each others, the
ACG would have a large number of nodes and that makes its size inecient.
The relevance of an attribute depends on the diversity of its values within the
service interaction log. A threshold is set to ignore attributes having low diversity
meaning that most of the messages have the same value for the same attribute.
In some other cases, messages of iterative service calls (loops in a process)
can occur in the log. The sequences of correlation conditions referring to a same
process model may have various lengths as each execution introduces a variable
number of loops, and thus includes a variable number of repetitive subsequences
in the ACG graph. Therefore, such sequences are not aggregated together and
lead to the identification of several processes.
Also, the approach works well if the log contains correlator attributes and
every two consecutive messages belonging to a same process instance are actually
correlated using some of those correlator attributes. Therefore, if the process
is larger, it is more likely to have broken correlation chains. This is the case
when two consecutive messages belonging to a same process instance are not
correlated using any pair of correlator attributes. In such cases, a large process
will be fragmented in the ACG as multiple smaller processes.
Finally, noise in service logs affects the result of correlation discovery as widely
observed [5] [6]. Real-world logs are imperfect, i.e., they are incomplete (corre-
spond to a subset of possible execution) and noisy (e.g., do not record some
messages). A known approach to deal with noise in logs is to use a frequency
threshold to filter noisy data [11]. In previous work [5], we have presented a
quantitative approach for estimating a noise threshold used to filter noise from
service logs. In this paper, we assume that the log is free of noise, or has been
cleaned from noise in a pre-processing step.
