interested in Deep Computing technology invariably ask one question,
according to Rick Lawrence, Manager, Parallel Applications Group,
Watson. "What they want to know is whether our algorithms and
fast computers can solve a problem that affects their bottom line,"
his group has shifted its focus over the past year from developing
new algorithms for the SP parallel computer to the larger issue
of how to use the technology to solve important problems in customer
relationship management and financial modeling.
This kind of
Deep Computing problem is known as datamining. It has long been
recognized that buried in the terabytes of everyday business data-credit
card transactions, grocery purchases, loan applications and stock
trades-are nuggets of valuable information. In the past, the challenges
of analyzing such huge databases have forced businesses to view
their customers through the simplifying lens of averages. Deep Computing
can move business beyond averages and toward individualized treatment
England, just outside London, the Safeway UK supermarket chain is
experimenting with a Deep Computing system that takes the hassle
out of grocery shopping. A palm-sized electronic organizer lets
customers compose a new shopping order starting with an automatically
generated list of all recent purchases. "You can build your
shopping list any place you can take the device," says Marisa
Viveros, Manager, Emerging Database Applications. For placing an
order, the device is connected via modem to a Safeway server that
processes the order and then transmits it to the Basingstoke store.
A Safeway employee gathers the groceries. The customer just stops
by and picks them up.
For this project,
the Research team focused on human-computer interaction technology,
scalable server infrastructure to support large numbers of customers
and personalized content. "Everything you display on that small
screen has to be relevant to the customer," Viveros notes.
The Deep Computing
aspect of this system is a feature known as personalized product
recommendation. IBM researchers have used a variety of datamining
techniques to look for shopping patterns among the spending histories
of 20,000 Safeway customers. "Not long ago, all customers received
essentially identical marketing messages," says Lawrence. "More
recently, market segmentation techniques have directed different
messages to different groups based on common interests. Now we’ve
reached the third phase, where we can tailor the message to the
to this third phase is ‘collaborative filtering’. "We’re using
datamining clustering to find groups of people who, in some sense
of their shopping behavior, are similar," explains Viveros.
"Then we find the most popular products they’ve bought, and
we use that as input to another level of filtering called content-based
filtering. Finally, we assign a score to each of these products
based on a matching algorithm."
The result of
all this Deep Computing is a list on the handheld device of 10 products
the customer has never purchased but is likely to enjoy. This is
meant to counter the possibility that when customers view the store
through the display of a palm-sized organizer they will stop making
serendipitous discoveries and impulse purchases.
In the United States, a large provider of consumer information also
depends on the datamining capabilities of Deep Computing. The company
wants to offer improved analysis that will enable financial
to achieve better targeting, and hence higher response rates, on
the two billion offers for credit cards that are mailed each year.
"We were given about a quarter of a terabyte of data, which
we loaded into IBM’s Universal Data Base on a 24-node SP at the
IBM Teraplex Integration Center-a demonstration facility for data
warehousing-in Poughkeepsie, New York," Lawrence says. "And
this represents only a small fraction of this company’s data."
filtered and aggregated, the data is reduced to a form that a parallel
computer can handle to discover how people actually use credit.
One of the most useful results identified a group of people
who are usually rejected for credit but who in fact are good credit
risks-a virtually untapped market.
datamining run on such huge databases with a single-processor system
used to take upwards of eight hours. "This is a fundamental
bottleneck in an analyst’s ability to acquire insight into this
data," Viveros says. Using powerful SP2 parallel computers,
the same analysis can be performed in half an hour. Which means
that Deep Computing can give decision makers plenty of time for
what is even more important, deep thinking.