Database Reference
In-Depth Information
Rule II: Learning by direct interaction is more important than analyzing purely
historical data.
Clearly, knowing whether a customer bought milk 5 years ago, and in what
combination, is less important than the information about his/her response to the
milk offer in the current session. And knowing what moves a chess player made
2 years ago is much less important than understanding what tactics he/she is using in
the present game.
1.5 Combining Offline and Online Analysis
Despite a trend toward realtime analytics, we have seen that both classic and adaptive
analytics methods have their pros and cons. Ultimately, it is futile to try to trade one
off against the other - both are necessary. Fortunately, they complement each other
perfectly: historical data can be used with offline methods to create the initial analysis
model so that the online system is not starting from a blank slate. Once the online
system is operational, the analysis model is modified adaptively in real time. Offline
analytics can still be useful when the online system is running, for integrating external
transactions which cannot be communicated to the system online.
Once again, chess can offer us a useful example here: an offline chess player
only learns by replaying games from chess topics. By contrast, an online chess
player only ever plays against living opponents. A combination of the two is ideal:
replaying and learning from other people's games and at the same time keeping up
with the practice.
To summarize,
Rule III: Offline and online learning complement each other organically.
For example, the recommendation engine of Sect.
12.3
, the prudsys RDE,
always combines both types of analytics.
1.6 Methodical Remarks
Before embarking on the actual subject of recommendation engines, we'll begin
with a few preliminary remarks on methodology.
Our principle, wherever possible, is to reduce a complex problem to simple basic
assumptions and then to address it in mathematical terms as fully as possible. In
other words, rather than tackling a problem in its most complex form and reaching
only vague conclusions, it is better to solve the simplified problem rigorously. After
that, it may be possible to use the knowledge obtained to solve the problem for more
