IEEE Technology and Society Magazine - June 2019 - 58

The optimization and run-time adaptation of query
execution is a component suggested to facilitate fast
response times during extreme-scale analytics across
different computing platforms, as well as the execution
for Big Data processing workflows (see Figure 1). Data
analytics is not necessarily performed on a single platform. Parts of the processing could be pushed to the
input sensor level, e.g., collecting simulation data, while
more computationally intensive operations, such as
population cell simulations, could be executed in one or
more (potentially distributed) Big Data platforms, or
within clusters (e.g. GPUs) of a supercomputer. Even
within a single supercomputer, one often finds different
available clusters, with different hardware and processing capabilities. In case of (machine learning) operators
that have available implementations in different platforms, it is desirable to minimize the data analysts'
involvement in the specification of the platform on
which these operators must be executed. Hence, optimization and runtime adaptation technology [8] is foreseen to automate the optimal allocation of resources
on-the-fly for the execution of these operations.
Distributed complex event forecasting adds up as a key
enabler component for the recommended architecture.
The streaming input is constantly matched against a set of
event patterns, i.e., arbitrarily complex combinations of
time-stamped pieces of information. An event pattern can
either be fully matched against the streaming data, in
which case events are detected, or partially matched, in
which case events are forecast with various degrees of
certainty. The latter usually stems from stochastic models
of future behavior [9], embedded into the event processing
loop, which estimate the likelihood of a full match, i.e., the
actual occurrence of a particular complex event.
Notably, "forecasting" in this context is not to be
confused with "predicting," typically used to refer to
machine learning models classifying previously unseen
instances. From a methodological standpoint, it also
differs from time-series forecasting. Complex event forecasting combines symbolic and numerical streams for
foreseeing the occurrence of any type of situation that
may be defined as an event, based on combinations of
other similar events and contextual knowledge.
Given that the input consists of a multitude of data
streams, events may correlate sub-events across many different streams, with different attributes and time granularities. Therefore, a highly-expressive event pattern
specification language, capable of capturing complex relations between events, is necessary. Moreover, the actual
patterns of what constitutes an event of interest are often
not known in advance, and even if they are, event patterns need to be frequently updated to cope with the drifting nature of streaming data. Therefore, the required
algorithmic apparatus must incorporate machine learning

58

techniques for learning and revising complex event patterns from data, which is another role of the component
of distributed learning of event patterns (see Figure 1).
What is necessary here is highly-expressive, declarative
event pattern specification formalisms, which combine
first-order logic, probability theory [10], [11] and automata
theory [12], [13]. This is due to the fact that such formalisms have a number of key advantages: i) they are capable
of expressing arbitrarily complex relations and constraints
between events; ii) they can be used for event forecasting,
offering support for robust temporal reasoning; and iii)
they offer direct connections to machine learning techniques for refining event patterns, or learning them from
scratch, via tools and methods from the field of Statistical
Relational Learning [14]. Notably, all existing techniques
for complex event detection and forecasting, as well as
machine learning techniques for the automatic extraction
of event patterns from data, need to be extended to work
in a highly-distributed, streaming setting.
Complex event forecasting contributes significantly to
the effectiveness of interactive analytics. The idea is to
use an event-based methodology to model time-consuming, computationally-demanding operations, and then
use event forecasting to derive good approximations of
the operations' outcomes, without the need to wait for
the actual operations to terminate. This will save time,
effort and resources (computing power, bandwidth), by
enabling data analysts get rapid responses from expensive operations, thus identifying undesired outcomes and
poorly performing configurations of particular data analysis approaches and facilitating the exploration of more
promising options. Equally beneficial in speeding-up
time-consuming computational processes to foster interactivity, is to use specialized domain knowledge for the
problem at hand, which allows simplification of the computational process by, e.g., pruning large parts of a
search space for some difficult optimization task, which
are known beforehand to be pointless to search.

Virtual Laboratory for Studying Tumor
Growth and Evolution
The introduced extreme-scale analytics architecture is a
key enabler to provide a "virtual laboratory" for studying
tumor growth and evolution. It can support the goal of
using in-silico models of cell systems found in in-vivo
tumors, to facilitate the design, testing, and optimization of cancer treatments based on combinations of different drugs.
In-silico simulations of such phenomena require modeling multi-scale processes, thus bridging the gap
between different levels of description and connecting
events that occur at different scales. As an example of
such correlated multi-scale phenomena, consider a DNA
mutation that alters the function of a protein, which leads

IEEE TECHNOLOGY AND SOCIETY MAGAZINE

∕

JUNE 2019
IEEE Technology and Society Magazine - June 2019

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - June 2019
