IEEE Technology and Society Magazine - June 2019 - 57

extracted data synopses and data streams are forwarded
to a number of recommended components responsible
for interactive analytics and operational proactivity.
Accounting for the requirements in the "Technological
Challenges" section above involves four main components
for: a) distributed machine learning and data mining; b)
data processing and workflow specification; c) distributed
complex event forecasting; and d) optimization and runtime adaptation of Big Data processing workflows. We
examine the operation of each component in turn, below.
The distributed machine learning and data mining
component is used to incorporate a wide variety of tools
for building various types of predictive models from data.
These tools can rely on existing algorithms for data analysis [6], [7], which need to be customized to work in a distributed, streaming setting, and optimized to cope with
the immense data volumes. The machine learning algorithms are utilized to constantly mine the streaming input
to generate, maintain, and update a rich variety of
diverse, high-quality models and model ensembles.
These models can be used, at any time, to analyze the
data that stream in. At the same time, these algorithms
may be used as a library for interactive data analytics, for
the data analyst to experiment with compressed data synopses. This will enable the data analyst to, e.g., try different parameter configurations per model, different feature

sets on existing models, or even generate new models onthe-fly from representative data views, thus gaining early
insights on the properties of the data she's dealing with.
The process of interactive data analytics is not supposed to hinder the actual knowledge extraction from
the streaming data: for instance, the life scientist should
be able to explore data stream synopses, as the
machine learning algorithms seamlessly integrate new
regularities into the models, by continuously updating
them from the entire stream of data.
The data processing and workflow specification component is added to offer additional support for interactive data analytics, via a rich variety of data processing
functions, as well as the ability to plug in new custom
functions, or new data analytics algorithms (see Figure 1).
Moreover, it allows the composition of data analytics
pipelines from existing models and data processing
functions. Interactive data analytics must be supported
by graphical interfaces for the specification of data processing workflows, thus enabling the domain expert, in
cooperation with the data analyst, to take the most out
of the platform with minimum coding effort. The vision
is to involve non-data scientists into fast loops of interactive, exploratory data analysis, where domain experts
and data analysts may gain rapid insights on properties
of the data "at a glance."

Applications
Interactive Analytics
Data Processing and
Workflow Design
Facilities
Source 1
Distributed
Streams

SDE

Distributed Learning
and Data Mining
Operational
Proactivity

Source N
Data Summarization

Distributed Complex
Event Forecasting

Optimization and
Runtime Adaptation
Pluggability
FIGURE 1. Interactive Extreme-Scale Analytics - The Big Picture. "Applications" emphasize the applicability of the approach to several
domains, while "Pluggability" refers to component modularity.

JUNE 2019

∕

IEEE TECHNOLOGY AND SOCIETY MAGAZINE

57
IEEE Technology and Society Magazine - June 2019

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