Database Reference
In-Depth Information
The significance of the first disadvantage, more complex theory, is often
completely underestimated. Developing an adaptive algorithm for an existing task
is not usually a problem. People tend to assume that the feedback loop will fix
any glitches, but they are wrong. Successfully developing adaptive methods, in
theory
and
in practice, is an art. For anyone who has ever come into contact with the
theory of adaptive error estimators of differential equations, most conventional
FEM solvers seem almost like light relief in comparison! The same is true of
realtime analytics methods. As we will see in this topic, their whole philosophy is
far more complicated than that of conventional data mining approaches.
Incidentally, the example of light waves we looked at earlier can also be used
quite effectively to illustrate the problem of developing powerful adaptive systems.
Simply getting a sign wrong (even just for a moment) in the third or fourth
Maxwell's equation would cause the entire electromagnetic wave literally to
collapse. It is not for nothing that physicists are constantly delighted by the
“beauty” of Maxwell's equations.
Philosophically, one could argue that the greater capability and robustness of
adaptive behavior comes at the cost of a significantly increased workload in terms
of theoretical and practical preparation. And yet it is worth it: once their develop-
ment is complete, the practical advantages of adaptive realtime systems become
abundantly clear.
The second disadvantage, restricted method classes, is related to the first. It is
not merely difficult to design conventional data mining methods adaptively; in
some cases, it is downright impossible. It is a fuzzy boundary: any data mining
method can be made adaptive one way or another, but fundamental features of
the method, such as convergence or scalability, may be lost. These losses have to be
weighed up and checked in each individual case.
The third disadvantage appears self-evident: realtime analytics systems need a
direct feedback loop; otherwise they cannot be used. In many areas, such as
product placement in supermarkets for cross-selling or the mailing of brochures
in optimized direct mailings, no such loop exists. There is nothing to be done about
this - other than wait. And waiting helps: the introduction of new technologies is
constantly extending the potential applications for technologies with realtime
capability. In supermarkets, these include in-store devices such as customer
terminals, voucher dispensers, or electronic price tags, which are currently revolu-
tionizing high street retailing. But online and mobile sales channels too offer
excellent feedback possibilities. The trend is being reinforced by a general move
within business IT infrastructure toward service orientation (SOA, Web 2.0, etc.).
If we look at classic and adaptive analytics methods, we can see a general shift in
the understanding of analytics methods. Until recently,
Rule I: The larger the available data set, the better the analysis results.
In statistical terms, that is still true of course. But increasingly, it is also the
case that
