1.Machine learning isn’t new. Machine learning, a method for transforming data into practical knowledge through the application of computer programs that automate the process of statistical modeling and data mining, has been around for a while. A lot of it is old statistical techniques rebranded to sell new software to run these techniques.
For instance, a week or so ago, I got a pitch from a company that noted it was using linear regression. Is linear regression some cool new AI feature? No, it’s a statistical technique that dates back to the 1800s, when the “machine” was your brain. What is new is our ability to combine data, algorithms and processing power. IBM’s Watson is a great example of applying decades-old algorithms to new “datasets” like Wikipedia. Machine learning is also based on two factors, an algorithm and data. IBM Watson’s algorithms were developed in the 1970s, but it didn’t get the data it needed to be effective until the 1990s or the 2000s.
2. Machine learning has practical applications. Even though ML has been around for a while, there’s a reason why we’re suddenly hearing a lot about it: There’s an abundance of data that didn’t exist even 10 years ago. That data will grow exponentially as IoT takes off.
Think about it: Soon, when a consumer gets up at 10 PM to fetch a beer while watching Game of Thrones, he leaves a data trail. His fitness tracker knows how many steps he took to the get the fridge, his smart refrigerator will know he’s down to his last beer and tell Alexa to order a fresh six pack, and Google will know his search behavior on his mobile phone. ML has the ability to find connections and patterns in that data that could provide new revenue opportunities and increase customer retention, among other benefits.
The other part of the equation is that there is now more computer processing power available than ever before. Using the cloud, companies have a nearly infinite amount of processing power at a price point that has never been lower.
3. All of this is pointless if you don’t have good data. What marketers don’t see as part of the software demonstration is the effort required to build the dataset that is the input into a ML algorithm. Building these datasets requires time, technical skill and business knowledge. If you want a ML algorithm to have a fighting chance of identifying patterns in your data that led to great insights, you need to assemble the right data sets in the right way. The marketing world needs better data, not better algorithms.
Temper Your Expectations
Marketers need to be skeptical about vendors’ claims. Ask them about their algorithms and terms like linear regression and pinpoint whether what they’re selling is really new or just repackaged algorithms that they’re calling machine learning.
Despite such sales bluster, the hype over ML is largely justified; you really can make huge strides by running a deep analysis of your data. But that assumes that you’ve already collected and assembled your data correctly and the information is solid.
Since most marketers are still stumbling around this step, their skepticism over ML is justified. But, their optimism about ML is also justified.
Many marketers are sitting atop a mountain of data about their consumers. Even in the face of the hype, marketer’s can still be optimistic about ML, just add a dash of pragmatism to go with the optimism.
Think deeply about your business. Identify the few specific – not general – problems where ML will help you make gains. After this, build a team and a process to build quality datasets, experiment with different ML methodologies, and learn what works for your business. This simple formula will ensure your expenditures on ML pay off.