IEEE Technology and Society Magazine - June 2019 - 56

rapid responses. They can also provide support on how
time-consuming computations may be broken down into
smaller chunks in an optimal manner, based on the specifications of the underlying computational resources, and
how results from such smaller chunks may be presented
to the life scientist in an incremental fashion.

Challenge 2
Big Data analytics tools mine past data to extract patterns conveying insights into what has happened, and
then project those patterns into the future to make sense
of the fresh simulation data that stream in. This permits
only the detection of such patterns, which is often inadequate. In order to mitigate risks, capitalize on opportunities, and allow for proactive decision making, predictive
analytics tools, enabling forecasts of future events of
interest, are required. Consider, for instance, the ability
to forecast the emergence of resistant cells via the detection of small changes in physiological or molecular markers. Interactive analytics, therefore, should be guided by
complex event forecasting. The ability to forecast, as
early as possible, a good approximation of the outcome
of a time-consuming and resource-demanding computational task also allows the rapid identification of undesired outcomes and saves a valuable amount of time,
effort, and computational resources, which would otherwise be spent in vain. Consider, for example, the possibility of forecasting the outcome of a complex multicellular
system simulation for tumor evolution, without the need
to wait for the simulation to be completed.

Requirement 2
Addressing this challenge requires cutting-edge techniques combining distributed, online machine learning
and complex event forecasting. Algorithms should incorporate, in real time, new knowledge that streams into
ever-changing, ever-adapting, but highly robust and accurate models, while also being able to make trustworthy
forecasts of future events of interest, fostering proactive
decision-making. Additionally, a library of distributed,
highly scalable, and optimized machine learning techniques, based on deep analysis of massive amounts of
historical data, allows the discovery of valuable patterns
of past behavior, which are subsequently subject to
change and adaptation via online revision techniques.

Challenge 3
Allowing the data analyst to gain rapid insights on the
characteristics of the massive input lays the ground for a
synergy between domain experts, human analysts, data
analytics algorithms, and computing infrastructures.
This is in contrast to the current state of affairs, where
algorithms and infrastructures are treated as black
boxes. Such a synergy could result in efficiently handling

problems, which are currently beyond the abilities of
contemporary information and computer technology. As
a simple example of such a synergy consider a data analyst able to pre-attentively identify an informative feature
set over a multi-cellular population, using her prior
domain knowledge and some first insights on the data
properties, gained by exploring data synopses. Proposing this feature set to a machine learning algorithm may
ultimately improve the algorithm's performance faster
than would be possible without human intervention. In
turn, this affects the available computational resources,
which can be allocated elsewhere.

Requirement 3
Interactive data analytics should foster such synergies
between humans and algorithms towards the effective
solution of difficult problems.

Challenge 4
Domain experts and even data analysts often do not
possess the necessary programming skills to code, optimize, and debug data processing operations over Big
Data and HPC infrastructures.

Requirement 4
A visual tool for setting up the desired data processing
operations that would automatically translate the setting
to optimized code is highly desirable. In contrast to mainstream solutions, what is needed here is to design and
develop a flexible, pluggable, distributed software architecture that will be largely programmable and set up by
graphical data processing workflows. This will allow the
non-data scientist, such as the life scientist, to take the
most out of the underlying computational resources, by
interactively querying data-at-rest and data-at-motion and
experimenting with a multitude of otherwise highly
opaque models with complex parameter configurations.

Method, Architecture, and Apparatus
Overcoming the challenges outlined in the previous section revolves around integrating two core concepts:
interactive data analytics and operational proactivity
via complex event forecasting. Figure 1 illustrates a
possible architecture and outlines an algorithmic apparatus uniting the two concepts.
As shown in the figure, extreme-scale data streams are
continuously acquired from the various (cellular system
simulation) sources. A dedicated, lightweight engine relying on approximate query processing techniques is introduced to extract compact synopses out of these data
streams [3]-[5]. Synopses are thought of as a prerequisite
that allows for fast response times during exploration of
the massive, high-speed input (in Figure 1 this corresponds to Synopses Data Engine - SDE). Then, both the

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