Case study: How Salesforce Einstein Discovery showed that attendance at a top university is not
the main predictor of high earning power:
■
At Gartner's 2016 "BI Bake-Off" at the Data and Analytics Summit in Dallas, Texas, we gave
representatives of several modern BI and analytics platform vendors university and college
student demographic data, payroll data and a demo script. In addition to showcasing functional
differences across critical capabilities, we asked them to combine the datasets and derive
insights about which university graduates would have the most earning power 10 years after
graduation. Given the number of variables and combinations available to explore manually, the
representatives did what expert analysts typically do. They explored their own hypotheses rst.
In this case, it was the "usual suspects" of leading universities — "because going to Harvard
means you out-earn those going to state universities, right?" While there was a relationship in
the data between attendance at top universities and earning power, all missed the most
important driver, one that is not intuitive. The biggest indicator of students' future earning power
in the data was not their university. It was their parents' income, and secondarily whether they
completed their degrees. We cannot say precisely why this is. Is it due to work and study habits
learned at home from high-performing parents? Is it because wealthier parents can pay for their
children to nish college, even if that means it takes ve or six years? We can, however, say that
parental income was not a driver that the respondents knew to look for.
■
By contrast, although we gave all the vendors in the vendor exhibit hall the same dataset, only
Salesforce Einstein Discovery uncovered the main driver after just a few seconds of ingesting
the data, automatically analyzing it and generating a narrative about the results (see Figure 2).
How often do business people draw suboptimal conclusions from their data? How often do they
explore what they think are the key drivers or attributes of an outcome variable and stop when they
conrm their hypotheses? How many times might there be other more important factors affecting
the outcome variable that they have not thought to explore? This is the root of the challenge with
the current paradigm. The desire to overcome it will drive the transformational nature of the next
wave of market disruption, namely automation of all aspects of the analytics workow in order to
improve the accuracy and timeliness of advanced analysis (in light of the human context), remove
bias, and elevate the skills of more users to citizen data scientists.
Since automation will enable expert data scientists to focus on specialized problems and on
operationalizing and embedding enterprise-grade models into applications, only the most accurate
and signicant insights will be acted on by users. Expanded use of automation should also translate
into fewer errors from the bias inherent in manual exploration.
Page 8 of 36 Gartner, Inc. | G00326012
我的内容管理
展开
