It doesn’t take long nowadays before someone will raise the issue of machine learning (ML) and marketing. Here I try and explain some of the basic issues in respect of ML and marketing. Arguably the key one being machine learning is often used for prediction rather than explanation. A good ML model might predict excellently. Still you may not know why it has made the prediction. (This is known as it having a ‘black box’ quality). That you don’t necessarily understand what the model has done can be a non-concern or a devastating critique depending upon your aim.
Defining Machine Learning
It can be hard to know what people mean when discussing machine learning. This is especially true given all the acronyms involved.
Maybe one of the most confusing disctinctions is between ML (machine learning) and AI (Artificial Intelligence). These are sometimes used interchangably but it is helpful to think of AI as being wider than ML.
Machine Learning involves when a prediction is generated by an computer algorithm. AI is broader being anything that uses technology to simulate intelligence. Basically you could never run into an example of ML in the street but you might run into an example of AI in the street. (For example, you might meet Rutger Hauer or Alicia Vikander — although probably not Rutger Hauer. Indeed that reference just dates me a bit).
Still ML is part of AI and so for many managers, and even writers, the distinction isn’t always critical. A lot of times when AI is mentioned in the business world you could change it to machine learning or vice versa and the point would stand. As such I will discuss some common features here and feature discussions of both ML and AI. So what can we say of machine learning (ML) and marketing?
How will the emergence of ML change business? Well one thing that is clearly happening is that machine learning is becoming cheaper. At the same time human labor/activity isn’t. (Sadly although computers get more powerful every year human intelligence is on a much slower increase. Although human intelligence may well be noticably increasin. You should, mostly, feel good about your children). The implication of the changing economics is that tasks that can be automized will be. Basically getting a machine to do some things is increasing good value compared to using human labor. Machines are often more reliable too as they don’t get bored.
Content moderation is the sort of task that machines can do much more efficiently than humans. It is repetitive and follows some basic patterns. Although when using machine learning you should expect the occasional high profile — ‘how did that happen?’ — disaster. Sometimes the machine will still miss things that humans ‘just know’. Where the stakes are high it might make sense to have a human for a final check. (Do pay attention to what your self-driving car is doing.)
Cheaper Means More Prediction
As prediction gets cheaper we might see it intruding more upon our lives. Maybe soon you can expect to see your retailer tell you what you want. We already have recommendations, will these become stronger. “Go on, Go On, Go On…..”. (Father Ted references may not work in North America — https://www.youtube.com/watch?v=7w0ZyfkukUs). Will the retailer ever know what you want better than you do? For more on the busines implications of ML see the book Prediction Machines.
If the retailer is just sorting out our shopping for us this would be certainly conveinient. It does lead to new questions: how then can we deal with conflicts within the person? By which I mean things that I might want (cookies) but shouldn’t really have (at least not too many). On a positive note — I know I’m stretching a little — maybe it’ll help with more environmentally sensititive shopping. An algorithm could better factor in environment concerns than human shoppers confused whether it is more environmentally friendly to get a product made with fewer pesticides or less plastic.
Prediction Not Explanation
Traditionally statistical tests aim for explanation. If the day is sunnier we can expect to sell more ice cream. We can then argue for a specific connection between sunshine and ice cream. “With 1 hour more sun we might expect to sell 6 more ice creams.”
A lot of times managers want a prediction not really an explanation. Here I mean they don’t want to know the abstract relationship between sun and sales. Instead they want: “How many ice creams will we sell today given everything that is happening? I don’t care what is driving it. I just want a number” The point being in a complex world lots of things are going on. You never get sunshine without other things that also impact ice cream sales. For example, temperature, day of the week, is the day in school summer holidays?, etc… Many factors determine consumer responses. In some ways it makes more sense for managers to make a prediction based upon all elements of the context. Managers might not need detailed explanations if the predictions work. They just need to know how many ice creams to get out of the freezer at the back.
Predicting The Present And The Past
It is worth clarifying that when we speak of predicting we don’t just mean predicting the future. Looking at the future can obviously be important. Future sales is likely somthing a manger wants to predict. Still predicting means estimating anything you don’t know.
So you may give a machine learning algorithm a picture of a squirrel. It will predict if there is a squirrel in the picture. It does this given its past experience of pictures of squirrels. You can get a % estimate for the chance of the picture containing a squirrel. If the machine is pretty confident that it is a squirrel that works for decisions that relate to whether a squirrel was present. Okay, I can’t think of why you want to know about squirrel presence but I just wanted to show the squirrel picture below.
You can more easily see why an online advertiser might want to look at what leads to clicks. This might be a picture of anything: the product, a person using the product, the product in a house, maybe an unrelated picture leads to clicks. (Is it pictures of squirrels that leads to greater interest? Of course it is. Ask our beagle — the answer is always squirrels.)
There is a Silicon Valley bit on image prediction that captures this well. One character designs an app that idenitifes food. (Warning the language is a bit much for some people and I’m not 100% sure how acceptable the portrayal is: https://www.youtube.com/watch?v=ACmydtFDTGs.)
Bias, In The Statistical Sense, Can Be Acceptable
Note it is perfectly possible for an algorithm to have bias (in the statistical sense) and still be useful. Your algorithm might give too much credit to sunshine and not enough to temperature in predicting ice cream sales. Still if the prediction overall remains pretty good the manager might not really care.
What matters is the quality of its outputs not the validity of its inner workings. Indeed, with black box models you won’t really know what is going on in the machine’s ‘head’ anyhow. As such, the black box quality means it is hard to know what is happening never mind whether the inner workings are valid.
A New Way Of Thinking
Using Machine Learning models requires a different way of thinking. It isn’t worse or better. It depends what you are trying to do. If you are explaining the impact of a single factor on another single factor machine learning isn’t likely to be for you. If you want a useful prediction you can act upon machine learning may well be the approach you want. Given this it is easy to see why managers with practical aims might embrace machine learning before academics, who require more detailed explaination.
Applying Machine Learning: Machine Learning (ML) And Marketing
The rise of machine learning will impact many parts of marketing. We recently published a paper in the Journal of Retailing. This highlighted the implications for, unsuprisingly, retailing. Machine learning (ML) and marketing have a lot of evident crossover in application.
Perhaps the most hopeful idea in our piece is that machine learning has the potential to bring managers and academics closer together. In many ways machine learning driven assertions are less sweeping. They don’t generally posit universal connections between variables but make more modest, condext dependent predictions. I am hopeful that academics might learn to embrace context and see the value of discussions embeded in specific situations. For example, we might get away from the idea of there being a general number for the value of loss aversion (2x) and think more about the contexts when loss aversion is stronger or weaker.
Clearly algorithms are making a big difference to parts of marketing already, for instance programmatic media buying, see here.
Supervised Versus Unsupervised Methods
There are a number of different types of machine learning. The two terms you are most likely to encounter are supervised and unsupervised learning. In many ways the distinction is pretty intutitive.
You may think of unsupervised learning as something you do when you do not have a perfectly defined end goal. This is more of a ‘fishing trip’ — exploratory analysis. Imagine a situation where you are seeking to find interesting things from the data. The analysis might show that two factors relate to each other. You weren’t explictly trying to find that out but it turned up and was quite interesting. In unsupervised learning you take the data and you find out what secrets it has.
Supervised learning is different in that you know what you want to find. For example, you want to predict what factors lead to a sale. You supply a bunch of training data which is labelled whether a sale was made or not. Basically the data contains lots of things that might relate to a sale as well as a record of whether a sale was made to create a model. The supervised element is that you supply what you want it to find — e.g., the record of the sale. You then use the model you developed on the training data to predict the sales you will have made in new data. (Of course you might not actually know whether the sale was made in the new data).
Seeing The Value Of New Methods
One challenge for managers is just to see the value of the new techniques. There is a great story of the UK government largely abandoning AI research convinced it would never work.
The blog post above again raises the problem of the black box. It can be hard to understand what the algorithm is doing. Some solutions seem pretty simply though. Imagine your facial recognition algorithm has been trained on caucasians. It therefore recognizes them better than other people. The solution seems quite obvious, you need to train the algorithm on a more diverse population. (Clearly there are issues with more challenging solutions but why not solve the easy ones quickly at least).
For more on AI see here.
Machine Learning (ML) And Marketing
It is possible to image a world of individual targeting of offers and messaging driven entirely by machines. (We are moving in that direction but I don’t think anyone can truly say they are there yet).
If you know what any individual likes you can tailor your marketing precisely to them. Again this leads to interesting questions about what the marketer actually knows. If the marketer sees that a certain consumer likes images with the color blue, the marketers doesn’t know why. That said, maybe the marketer is happy to just show the consumer blue images and not worry too much about why. (Of course this only works while the relationship between color and happiness stays the same. If the relationship changes — the consumer starts to prefer red images — you won’t really be able to understand the change given you never really understood the reasons for the blue preference in the first place.)
For more on targeting (specifically personalization) and AI see here.
There is of course a downside to using machine learning. (There always is I guess). Specifically the black box quality. A major problem with prediction when you don’t care ‘why’ is that the ‘why’ can turn out to be very important.
ML models that are based upon data from the world can codify any exsiting problems into the algroithm. Imagine you want to automate who gets loans. It is quite possible to create a model that looks at previous applications (approved and non-approved) and mimics the process. It can become very good at predicting who would get loans in the current system. You then impliment this system. You save money given you have cut out human labour but nothing changes in the process. That is a problem if the system wasn’t fair in the first place. What is more you now have a black box machine that has bias hidden in its internal workings that no one can point to. It seems impartial because it is a machine and doesn’t harbor any malice but it isn’t really.
We do need to be careful about taking the problems we have as a society and baking them into code that we don’t understand and can’t easily address.
Bad Does Not Show Progress Has Not Been Made Given The Past Was Very Bad
That said, I do have a personal plea. It is an error I see being made a lot. Just because a problem currently exists does not mean that progress hasn’t been made or, worse still, that things are getting worse. When you see a problem things may be getting worse. They may also be staying the same. They may also be getting better. (This is often too slowly and we may need to understand that they will only continue to get better if people keep attention on the problem).
My point is that it is very possible that big data will increase inequality. Still when we say things are getting worse it is worth remembering that things were always far from perfect. Machines have problems but people do too and always have had. We have much more evidence of humans behaving badly than machines.
To understand if machines are making the world worse we need to recognize where we are starting from. Maybe things are getting worse, or maybe they are getting better. That things are getting better seems to fit the facts to my mind see here and here. Of course, different contexts may have different answers to the question. Lets use evidence either way.
I really want to push back on progressives who take the attitude that any injustice shows we haven’t made any progress. That is like saying we all die so medicine doesn’t help at all. Machines certainly won’t solve all our problems. Though if things aren’t perfect in a few years (and they won’t be) it won’t necesarily be the machine’s fault. Humans were quite able to mess things up long before the computer was invented. Machines may help us improve the world but don’t expect miracles.
A problem to bear in mind is consumer reaction when they have to interact with some sort of bot. Evidence suggests that consumers aren’t that keen, see here.
Conclusion: Machine Learning (ML) And Marketing
I hope you have found this discussion of machine learning (ML) and marketing interesting. This is clearly an area where greater clarity is needed including in this piece. I would especially welcome suggestions on improvements and about things that didn’t make sense to you.