However, instead of more data, analysts need specific data. “Ninety-five percent of users are looking for information that’s relevant to their jobs and supports decision making and improved performance,” Johnson says. Instead of focusing on more data, organizations must consider the business users to determine not just what data they need to access, but how it’s presented.
“Providing access to every bit of information and in multiple formats can be overwhelming and actually stifle adoption,” Johnson says. “Instead, find out what’s important to them and how you can present that information to them in the simplest possible format.”
Myth 3: Analytics removes human bias
The way automated systems perform is not supposed to be biased. But technology is built by humans, so eliminating all bias is nearly impossible. Some think analytics and machine learning remove human bias.
“Unfortunately, this is not true at all,” says Mike Mason, head of technology at ThoughtWorks, a global technology consultancy. “Algorithms and analytics are tuned using ‘training data’ and will reproduce whatever characteristics that training data has.”
In some cases, this will introduce a benign bias into analytics results; in other cases, a more severe bias, Mason says. “Just because ‘the algorithm said so’ doesn't mean the answer is fair or useful,” he says.
Myth 4: The best algorithm always wins
It actually turns out that with enough data, “sometimes the algorithm doesn't matter,” Mason says. In an IEEE article he cites, "The Unreasonable Effectiveness of Data," Google engineers argue that simple statistical models, coupled with extremely large amounts of data, achieve better results than an "intellectually superior" model containing lots of features and summarizations.
“In some cases, just crunching a bigger pile of data will achieve the best results,” Mason says.
Myth 5: Algorithms are fail safe
People inherent trust statistical models and algorithms to a high degree, and as organizations build their analytics programs they increasingly rely on sophisticated models to support decision making, Johnson says.
“Because people don’t understand the models, algorithms, and other advanced data science practices, they place their trust in them,” Johnson says. Users don’t feel like they have the knowledge to challenge the models, so instead they must trust the “smart people” who built them, he says.
“For the last 50 to 60 years we’ve heard AI will take over in 20 years, and we’ll continue to hear people say that,” Johnson says. “There is still a lot of ground to cover before we can blatantly trust machine learning and the results. Until then, we need to challenge the people who build algorithms and models to explain how answers are reached. It’s not that we can’t depend on the results; it’s that we need transparency so that we can trust and verify analytics.”
Sign up for Computerworld eNewsletters.