I don’t mean to alarm you, but the Marketing Singularity is near. It is the point in time when artificial intelligence (Robots) will be able to handle marketing as well as (or better than) any human marketer can.
What will give robots a marketing edge over us humans is Machine Learning: an exciting and relatively new subset of computer science in which complex algorithms and powerful computers combine to enable robots to instantly “learn” and adapt to new, untested scenarios.
For a great analogy and breakdown of what Machine Learning is, you can read this thread on Quora:
In it, Pararth Shah of Stanford lays out an analogy of shopping for oranges from street vendors. The objective in this scenario is to pick the best oranges for the best price. Here is how a Machine Learning program could help:
You take a randomly selected specimen of mangoes from the market (training data), make a table of all the physical characteristics of each mango, like color, size, shape, grown in which part of the country, sold by which vendor, etc (features), along with the sweetness, juiciness, ripeness of that mango (output variables). You feed this data to the machine learning algorithm (classification/regression), and it learns a model of the correlation between an average mango’s physical characteristics, and its quality.
Next time you go to the market, you measure the characteristics of the mangoes on sale (test data) and feed it to the ML algorithm. It will use the model computed earlier to predict which mangoes are sweet, ripe and/or juicy. The algorithm may internally use rules similar to the rules you manually wrote earlier (for eg, a decision tree), or it may use something more involved, but you don’t need to worry about that, to a large extent.
Voila, you can now shop for mangoes with great confidence, without worrying about the details of how to choose the best mangoes. And what’s more, you can make your algorithm improve over time (reinforcement learning), so that it will improve its accuracy as it reads more training data, and modifies itself when it makes a wrong prediction. But the best part is, you can use the same algorithm to train different models, one each for predicting the quality of apples, oranges, bananas, grapes, cherries and watermelons, and keep all your loved ones happy – Pararth Shah
Picking oranges is a rather simple task, and Pararth made it easy to understand how Machine Learning could help. But what complex tasks (the kind that human brains are needed for) can be made easier with Machine Learning? Here are a few examples:
From Bloomberg: The $10 Hedge Fund Supercomputer That’s Sweeping Wall Street
Braxton Mckee has written financial modeling software that uses machine learning and Amazon’s massive cloud-computing service to solve trillion-entry problems in only a few minutes, all for only $10.
From Science20: Machine Learning And Big Data Is Changing Sports
A great write-up on recent advances in sports analytics made possible through machine learning, including a research team’s model for identifying the key physical characteristics among valuable Australian Rules Football draft players.
Eric Malmi, a researcher at the University of Aalto has built a Machine Learning algorithm that identifies significant verses in rap songs, then writes new lines that rhyme in the same way and are on the same topic.
I don’t know how much use there is in the world for a Hedge Fund-guru, soccer-playing, beat-making robot, but these are all examples of how Machine Learning is being used to solve problems that, historically, only human brains were capable of.
There is, in my opinion, no field in science or the arts that is more human-focused than marketing. Marketing is, at its core, the practice of distilling human behaviors and emotions into actionable insights and reproducible results. Building an advertising campaign with a positive ROI is difficult for most humans, so how will robots do it?
How Can Machine Learning Be Used In Marketing?
Machine Learning is a relatively new field in data science, so unfortunately for me Barnes & Noble doesn’t carry any copies of Machine Learning for Dummies. To research Machine Learning in the marketing industry, I turned to the Google search engine – which if you’ve seen the new movie Ex Machina, you know it to be a Machine Learning robot in and of itself.
I was able to find a plethora of publicly available research papers on Machine Learning and how it can be used in marketing. They weren’t easy to read, but they were easy to find. If you are interested in this subject, I highly recommend printing out all these documents I link to and reading them, rather than reading my summary. This is some cool stuff, I swear!
Now, for the people who don’t want to read academic papers, here are three roles in marketing that Machine Learning robots can and will replace humans in:
1. Market Segmentation
Market Segmentation is a crucial part of any company’s marketing strategy. You need to identify your target consumer accurately before you ever think about creating and launching any advertising campaign, otherwise, you could be advertising the wrong way to the wrong people.
The paper I read about Machine Learning applications for market segmentation is called
Big Data-Driven Marketing: How machine learning outperforms marketers’ gut-feeling, and it was written by Pĺl Sundsřy, Johannes Bjelland, Asif M Iqbal, Alex Sandy Pentland, and Yves-Alexandre de Montjoye. You can read the paper here on Scribd.
This paper describes an experiment wherein Machine Learning was used to help a large phone company in Asia segment their existing users for a special promotion. The company wanted to send out a mass text to a select group of customers that were only using a prepaid plan; the text would offer a discount on a data plan. The test was set up as follows:
The phone company’s marketing team was asked to define their own segment for the promotion, using their “gut-feeling.” According to an IBM survey quoted in the paper, “80% of marketers report making such (market segmentation) decisions based on their ‘gut-feeling’”
- Independent of the phone company’s team, researchers developed a Machine Learning algorithm to analyze past customer actions and identify “behaviors of natural adopters,” or customers that bought the data plan without any promotional efforts. Then, using these behaviors, the algorithm would create its own segment to test against the human marketers’ gut-feeling.
Here are the “top 10 most useful features to classify natural adopters,” as identified by the algorithm:
“Total spending on data amongst close social graph neighbors” was the number 1 most useful feature, and that is something that human marketers were previously unable to calculate because of the sheer amount of data involved. It is a one-value feature derived from analyzing each individual user, comparing how physically close they are and how often they communicate, and plotting them on a “social graph.” Basically, it’s a number that shows how social this user is with their friends on their mobile phone. Using this complex feature, the researcher’s algorithm was able to outperform the human marketers.
Here are the results of the test, in terms of conversion rates:
- Human marketers’ segment conversion rate: 0.5%
- Robot marketer’s segment conversion rate: 6.42%
The robot’s campaign performed 13 times better! I would have been a little bit embarrassed if I were one of those human marketers…and scared for my job. This one test is pretty solid evidence for how valuable a Machine Learning algorithm can be for market segmentation, as the paper outlines in its conclusion:
The algorithm only relied on existing customer purchase data and did not require any new data, so it was much cheaper than the old, human method of taking surveys, analyzing results, testing a campaign, and repeating. Otherwise known as Trial and Error.
When using such an algorithm, segments can be much more accurate, and therefore, much smaller. This cuts down on spam, something marketers never want to do with existing customers.
In the “Discussion” section of this paper, the researchers theorize that an algorithm like the one they used can be used to help launch new products, which brings us to the next marketing role that robots will soon take over:
2. Predicting Consumer Preferences
When designing new products, companies might spend as much time and resources on figuring out what customers actually want as they do on actual research and development. Predicting how your potential customers will respond to your product is difficult…for humans, that is.
The paper I read on Machine Learning applications for predicting customer preferences is called Improving Preference Prediction Accuracy With Feature Learning, and it was written by Alex Burnap, Yi Ren, Honglak Lee, Richard Gonzalez, and Panos Y. Papalambros. You can read it here on Scribd.
The researchers behind this paper wanted to answer the question: How can we obtain high prediction accuracy in a consumer preference model using market purchase data? In simpler words, how can Machine Learning be used to turn old data on customer purchases into accurate predictions for future customer preferences? The “old data” part is important, because attaining new data is always time-consuming and money-consuming, which companies are always looking for ways to avoid.
The paper first describes the normal, human methods of predicting consumer preferences, which go something like this:
- Developing statistical models to analyze consumer surveys
- Creating “adaptive surveys,” which change as the subject is taking them in order to capture more significant data in a more efficient way
- Finding “covariates,” or significant variables that have comparatively high effects
That sounds time-consuming, right? The researchers behind this paper wanted to find a way to use Machine Learning to achieve the same results, but without gathering expensive, new data. Their solution was to design an algorithm that would “Learn’ features from the existing data that are more representative of the consumer’s decision-making process”. Or in other words, they wanted to train a robot that could use historical data to build better, more efficient models for predicting consumer preferences.
The test that the researchers performed in this paper concerned the auto industry. The only data that they fed into their Machine Learning algorithm was from three consumer reports from 2006 – all of which were several years old when this research was done. From this massive pool of car purchase data, the researchers singled out 212 different car models and 6556 unique consumers.
Then, the researchers utilized two methods of “feature learning,” or identifying attributes that are most significant in consumer preference. Those two methods are called “Sparse coding” and RBM, or “Restricted Boltzmann Machine.” Here is a graphic that explains the flow of data in an algorithm like this:
Now for the results. In the test, there was a control model made with the standard human methods, and two models made by robots. Here is how they performed:
- Human marketers’ reference model prediction accuracy: 71%
- Robots marketer’s reference model 1 prediction accuracy: 76.4%
- Robots marketer’s reference model 2 prediction accuracy: 81.4%
The robots win again! The researchers demonstrated two separate, successful methods of applying Machine Learning to predicting consumer preferences using existing data, and that is only the beginning. This paper claims to be the only example of a proven model like this so far, so it’s safe to assume that other researchers are improving on these methods right now. Humans have a lot of catching up to do.
But what if there were robots that are already replacing marketing jobs?
3. Targeted Display Advertisements
If you’re a digital marketer, hopefully, you’ve heard about Real-Time Bidding for targeted display ads. It is a recent advancement in marketing, one that was made possible by smart computer algorithms and massive computer banks. Google operates DFP, a massive RTB ad platform, as does Facebook with their Facebook Ad platform.
These ad platforms work like old media exchanges. With advertisers bidding to place their ads in specific places, only now there are millions of impressions to serve each minute and bidding takes place in real-time. Advertisers want to spend the least amount of money to reach the most amount of potential customers, so they want their campaigns to be well-targeted. That’s where Machine Learning comes in.
The paper I read on Machine Learning applications for targeted display advertisements is called Machine Learning for Targeted Display Advertising: Transfer Learning in action, and it was written by C. Perlich, B. Dalessandro, O. Stitleman, T. Raeder, and F. Provost. You can read the paper here on Scribd.
This paper is essentially a case study on how Media6Degrees (now called Dstillery) has been using Machine Learning to help companies target their display campaigns on RTB platforms. The challenge that Dstillery specializes in is reaching “prospective customers,” or, consumers that have had no interaction with a brand before. The goal of Dstillery’s complex algorithms is to build predictive models for campaigns instantly without any testing, which is the human method of targeting RTB ads.
For each campaign they run, the company wants to use all available data in order to make the best decision in targeting. So, they collect data specific to individual users, including:
- User browsing history, which is available through cookie tracking
- Bid requests from the RTB platforms
- Ad impressions
- Clicking on an ad
- Making a purchase on the brand’s site
- Take any other “brand action,” such as visiting the brand’s site
Here is a graphic that outlines the different stages that Media6Degree’s algorithm goes through for each campaign:
Remember, these robots are running through these stages 1000s of times per second, handling campaigns for hundreds of companies at the same time. I don’t think there’s any human marketers that can do that.
Media6Degrees routinely run control tests against their algorithm’s campaigns, which means that they take the same ads and run them without any targeting – a truly random test. Here is a chart showing the overwhelming success of their algorithms:
I hope that wasn’t too overwhelming. What this chart shows is a direct correlation between using Machine Learning algorithms and conversion rates of targeted display campaigns, but what is a better representation of this is Dstillery’s success. They have received $52 million in four rounds of funding, they now handle analytics for leading brands like Verizon, William-Sonomoa, and Orbitz, serving millions of ad impressions per day, and they only have ~100 humans working for the company. That is the power of Machine Learning in marketing, and it’s only beginning.
If you have any questions about Machine Learning, market segmentation, consumer preferences, targeted display, or anything else, reach out to us humans at Streetwise Studio.
*note: The robot image in the header is from Isaac Asimov’s Robot Visions