Hitmetrix - User behavior analytics & recording

What Machine Learning Can (and Can’t) Do

On first hearing, “machine learning” and “Artificial Intelligence” sound like technologies that will replace people. Computers will find people and sell them stuff they want, so who needs humans in marketing?

Well, it turns out that you do. Computers can do the scut work of counting, but only humans can truly say what counts. Marketers are not going to be replaced by automation, but they can make best use of it so long as they know what it can do and can’t do, just like any other tool.

Yes, Master…

Take note of what machine learning can and cannot do. You can program it to seek attributes, to count how many clicks a web site gets, “to learn from the data for a long time.” says Anil Kamath, Adobe Fellow and VP of technology at Adobe.

Machine learning can recognize campaign subject lines, tag images for visual search, analyze sentences, undertake real-time decision-making, power recommendation engines, and engage in real-time bidding, he pointed out.

Machine learning or AI is best applied either when there is a low yield in a business process or a large consumer surplus is generated from applying AI. While there are many functions where machine learning can be applied effectively, marketing, drug discovery or patient monitoring are sweet spots for machine learning.” says Aman Naimat, SVP of technology at Demandbase.

“[W]e should not apply machine learning to tasks where humans are very effective, like air traffic control at an airport.  If a task is already optimized, incorporating machine learning would not serve to maximize any return on investment.” he says.

Machine learning is “a buzz word mixed with AI and chatbots,” says Meghan Keaney Anderson, VP of marketing at Hubspot. “It’s really about a type of programming that looks for patterns in the data and tries to learn from past patterns.”

“By learning, [AI] becomes smarter in the process,” she adds.

“We’ve seen artificial intelligence and machine learning across the entirety of sales and marketing.” says Nipul Chokshi, VP of marketing at Lattice Engines, a specialist B2B solutions provider. Machine learning is great at spotting patterns, can highlight segments for targeted marketing, and spur demand acceleration further down the funnel. It should enable the marketer to deliver a smarter message, he said.

Avoiding Error

Before machine learning becomes effective, the machine — obviously — has to learn. Programmers will shovel terabytes of data into the hopper, all gladly digested by the learning algorithm. Ironically, the system is no less human than the people who built it. That means errors will be lurking. They will have to be screened out.

Take Lattice Engines. It offers a solution that looks for patterns in the data to identify who is likely to convert to a sale. “We use 80% [of the data] for our training model. “ Chokshi says. The remainder is set aside. “We use it to test the model to see if it could predict accurately,” he said, because the answer is already known.

“[I]f machine learning is king, data is his queen. If you don’t have enough, or rich, training data, no machine learning algorithm is going to work.” adds Naimat. “The best way to ensure quality of data is actually using more data. We often triangulate confidence in our data by getting different perspective on the same data from different sources. If lots of sources agree, then it’s more likely to be true.”

“For us, our data, especially analysis and other solutions—is used by the marketer right now.” says Adobe’s Kamath Data is used in the training and it’s possible to measure the validity, improve the model, and understand the differences that should be addressed. “Put in an extra layer to remove the outliers.” he said. If something is not right with the data, “we will see things suffer.” he says.

Quality is not entirely guaranteed, but quantity can make up for it, Anderson notes. “With small data sets, the impact of bad data is exponentially worse.” she said. “Whenever we see people fail, it is because (they) applied too small a data set.”

Keep a Human in the Loop

Machine learning is a co-pilot, not an autopilot. A person is needed to make judgment calls on the machine’s output.

“[This] is still a relatively new technology and, consequently, there is still room for error. There are a few ways we can put constraints on the problem.” Naimat says. “All AI systems should take human feedback or overrides and also provide justification for their actions. We have found that transparency of AI actions is an absolute requirement for building trust with its users.”

Machine learning can do things that are not humanly possible, like sort through years’ of data in minutes. Likewise, humans can do things that are not “machineley” possible. Does the pattern spotted by the machine make sense? “Or is it an anomaly and doesn’t make sense?” Anderson asks. When there are too many variables to count, let the computer count them. “A person has to do the second look.” she insists.

“People used to be skeptical of machine learning. They used to think of it as a black box.” Kamath says. Now machine learning is more sophisticated. It can be programmed to understand the customer journey, not just touch points.

Yet in all this, a human is in the loop, controlling the cycle and the data-to-machine learning experience. The human has the ability to change the algorithm. Adobe makes sure the model is visible, the controls are visible and the improvements are based on access to the data. The controls have to be visible, he said

Typically, machine learning will get you the same results, only faster. Take an analysis of an e-mail campaign marketing campaign. How many e-mails were opened? When? Those patterns are obvious to the user when looking at the output. But the system may also show unexpected patterns, like a particular kind of content that draws a larger than expected response, Anderson noted. “It may not be what you are looking for,” she says, but you may want to boost that content with additional ad dollars.

Usage brings proof, tempered by human judgment.

As Chokchi describes it, the system has to deliver information that sales reps can use to score more conversions, turning prospects into sales.  The algorithm can provide contextual information, so that the sales person does not have to waste time asking the prospect to describe the company’s systems and needs, he explained. And those prospects can be scored to show the likelihood of conversion.

An experienced sales rep can be a tough customer. “They bring in big deals and interesting prospects,” Chokshi says,but their work can be drawn out over weeks or months. So when an algorithm recommends a prospect, they are like “are you f—ng kidding me?” he said. “They are cynical and rightfully so.”

The Next Turn

Every company has a wish list for machine learning. Wishful thinking eventually becomes a wish that comes true.

In the near future, Lattice wants to provide a platform where it can take its own data, do its own scraping, and combine it with customer data. “The more data we can provide the higher quality recommendation we can make.” Chokshi says.

In the future, Adobe wants to see more integration between machine learning and  the customer experience, integrating data streams from web sites, apps and social media. “We got people to be good at channel (marketing) we need to look with a more holistic view.” Kamath says.

“Ultimately we are moving towards re-creating the intimacy that was lost by …the previous generation of marketing technology that produced spam but wasn’t effective.” Naimat says. “We want to create a unique personalized experience for buyers where everything is catered to their problems and communicates in their vocabulary… Not only we want AI and machine learning to provide insights, but in-fact be biased towards taking action directly rather than recommending actions.”

“I’m really interested in how it  (machine learning) allows us to better serve the customer.” Anderson says. Think about the chatbot married to the phone tree. “You end up with a negative experience,” she said.

A better alternative would be a system that reacts to the caller’s needs, but in a subtle way that is invisible to the customer, Anderson continues. “I don’t think we are there yet, but in a few years we will be.”

Total
0
Shares
Related Posts