IEEE Technology and Society Magazine - Fall 2014 - 38

its sub-events. A sub-event can be
described as event that comes from
sensors or other composite events.
Consider, for example, the recognition of a traffic incident in a road
segment (the composite event) and
the speed of the vehicles passing
that segment (e.g., simultaneous
deceleration sub-events).
Typically, event recognition systems operate on top of stream processing platforms [1]. This way,
complex events are defined by means
of expressive event recognition languages, and efficiently detected
using the optimized data processing
of stream processing platforms.
Event recognition systems have
to deal with various types of uncertainty, such as incomplete data
streams, erroneous data, and imperfect composite event definitions
[5]. For example, in traffic management, fixed sensors are often out of
order, inappropriately calibrated, or

minimize the performance overhead
of uncertainty reasoning, it is necessary to exploit distributed probabilistic reasoning techniques.
To allow for proactive decision
making, a framework for event forecasting is required, which is able to
deal with the volume and lack of
veracity experienced when relying
on Big Data. This framework should
indicate the probability of a forecast
event, as well as the probability of
when an event will happen; a probability distribution over the expected
event occurrence time can then be
provided. The basis of such a framework is "forward" event recognition
algorithms that are capable of recognizing incrementally composite
events, but incapable of dealing with
the lack of veracity [8].
The manual development of
composite event definitions is a
tedious, time-consuming, and errorprone process. Machine learning

Proactive computing requires
capabilities for forecasting,
real-time decision making,
and visual analytics.
inaccurate. To address this requirement, there is a need to develop a
framework for real-time event recognition able to deal with the inherent uncertainty of Big Data.
This framework should exhibit
declarative, formal (probabilistic)
semantics. To satisfy this requirement, we will build upon existing
frameworks combining probabilistic reasoning, such as Markov Logic
Networks [10], and symbolic methods. The starting point will be probabilistic extensions of the Event
Calculus [2], [5], a logic programming language for representing and
reasoning about events and their
effects. Probabilistic Event Calculi
facilitate the integration of domain
knowledge, such as traffic models,
and deal with the issue of uncertainty, both in the input data, and
the composite event definitions. To
38

This combination of techniques,
however, does not scale to Big Data.
In addition to learning the structure of a composite event, the confidence values/weights attached to
the definition can be learned from
data. Usually the tasks of structure
learning and weight learning are
separated. First the structure of an
event definition is learned, and then
the weights of the definition are estimated. Separating the two learning
tasks in this way, however, may lead
to suboptimal results, as the first
optimization step (structure learning) needs to make assumptions
about the weight values, which have
not yet been optimized.
To address these issues and avoid
the error-prone process of manual
composite event construction, the
methodology also consists of incremental learning techniques for successfully combining abduction with
induction in Big Data. Furthermore,
this should be complemented by
techniques for the simultaneous optimization of the numerical parameters of a composite event definition
(weights and numerical temporal
constraints) and its structure.

Scalable Processing

techniques may be used for the
acquisition of domain knowledge:
constructing and/or refining composite event definitions (expressing, e.g.,
traffic congestion) in dynamic and
evolving environments. A common
technique for learning the structure
of a composite event in a supervised
manner involves the use of Inductive
Logic Programming (ILP) (e.g. [17]).
ILP constructs theories that capture
exceptional cases in data streams.
This is particularly helpful in highly
imbalanced streams, such as those of
credit card fraud. On the other hand,
ILP does not handle numerical reasoning, such as comparing the timepoints of events emitted by vehicles,
which is quintessential in the representation of a composite event. In
the case of partial supervision, ILP is
used in combination with abduction
in order to learn an event definition.

The high velocity of incoming events
poses challenges both in terms of
computational resources and in
terms of communication resources.
Computational scalability issues are
addressed by distributing event-recognition tasks among multiple nodes
(see, e.g., [16]), while issues of communication scalability are addressed
by algorithms that perform as much
of the processing as possible on the
nodes where events are generated.
This ultimately reduces the amount
of data that is transferred between
nodes (see, e.g., [15], [9]).
In traffic management, a common task is counting the number
of vehicles traversing on a set of
paths, where some of the paths may
have shared locations (consider,
e.g., paths {A, B, C} and {A, D, B}).
A simplistic approach that does not
take uncertainty into consideration

IEEE TECHNOLOGY AND SOCIETY MAGAZINE

FALL 2014

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - Fall 2014