For almost any business or organization interested in using science and technology to provide a competitive edge it’s impossible not to have noticed the tidal wave of media reports on the impact and importance of Artificial Intelligence to just about anything – from issues of national security, such as cyber-security, drones, voter manipulation and fake news – to self-driving cars, smart-homes, algorithmic trading, and even to playing complicated games such as Go. Joining the tidal wave of media reports on AI has been an explosion of interest in the concept of “data science”, with Harvard Business Review calling data scientist “The Sexiest Job of the 21st Century”. This interest is societal, not just associated with scientists or technologists, as is in evidence with a quick look at the Google search statistics for these terms: Data science – 68,800,000 hits; Artificial Intelligence – 120,000,000 hits; Deep learning – 16,400,000 hits; Machine learning – 338,000,000 hits. In contrast, Business Intelligence, in some ways a much wider and more entrenched discipline in business, has 112,000,000 hits, while Evolutionary Computation – a sub-discipline of AI comparable to Machine Learning has only 564,000 hits. What this tells us is that there is a lot of general interest in AI, data science and machine learning. But is it all hype? Or is it really a paradigm changer, a turning point for business and society? Also, what should a CIO think of all this and do about it? What does a CIO need to know?
Here, I’d like to play the role of both advocate, and devil’s advocate, trying to cut through the hype and give, hopefully, a realistic appraisal of where we are and where we might be going. First, let’s talk about data science, being a scientist (physicist) myself, I need to first remove a “piedra en el zapato” about data science: All science is data science! In other words there isn’t any science worth the name that doesn’t use data. Science is about organizing data in order to understand phenomena. No data, no science. So, why do we use the term “data science” as if it was something different? Wikipedia says it is “an interdisciplinary field of scientific methods, processes, algorithms and systems to extract knowledge or insights from data in various forms, either structured or unstructured”. That doesn’t help much. Really, the chief difference comes from the quantity and type of data that is being used to study a system and, more importantly, what type of system the data comes from. It is this latter property that is the key to understanding why you might differentiate between science and data science.
Unlike data science, much of traditional science has been associated with understanding the physical world, where things are pretty simple – you drop a ball and it falls to the ground described by the Law of Gravity, which is a machine, an algorithm, that takes as input where the ball is now, and its velocity, and turns that into a prediction about where it will be at any time in the future. The physical world is like a machine and machines are simple and predictable, at least when compared to humans. We also fall according to the Law of Gravity, but we aren’t that simple, we do a lot of other interesting stuff. Although humans are very complex, we are definitely not unpredictable. If we were there would be no business, in fact there would be no human beings. We are, in fact, Complex Adaptive Systems, as is your business – that, I believe, is fundamentally a much more important label than machine learning or deep learning or data science. Complex means that, just like your business, there are many, many factors that influence what happens in it and to it and, adaptive means that, just like with your business, the future isn’t written in stone. Your best laid, most thought out plans can all be invalidated by the action of a competitor adapting better to prevailing macroeconomic conditions, for example, or by adopting new, superior technologies sooner via their own equally well thought out plan.
To describe Complex Adaptive Systems you need a lot of data – Deep data. You also need a different type of approach to modeling them, using AI and Machine Learning for example, rather than simple statistics or descriptive analytics. So, data science is really just applying the standard scientific method to Complex Adaptive Systems and this needs large, deep databases and advanced modeling techniques. And business, which is about human beings selling products and services to other human beings, either individually, B2C, or as organizations, B2B, is also a Complex Adaptive System.
Business is about human behavior and behavior is about making decisions and making decisions is about making predictions. That’s an important take home message. Everything a human being does, every human action, is associated with decisions: Do I buy a new car or not?; Do I stop smoking or not?; Do I invest in Hadoop or Cloudera?; Do I hire this Senior Data Analyst or not? A key fact is that every single human decision has behind it a prediction with respect to one or more cost-benefit measures. For instance, buying a new car you are predicting just how much the benefits – need for transportation, social credit, pleasure etc. – will outweigh the costs – price, reliability, financing scheme etc.
Human behavior and decision making – what we do – depends on who we are, how we think and how we feel. However, to describe who we are, how we think and what we do, requires enormously more data than that needed to describe what a machine is and what it does. We can predict pretty well what machines will do because each machine is designed to do only a very small number of things, and hopefully do them well. That’s not true with humans. We can do almost anything and our behavior reflects that. However, if you can predict behavior: Who will buy my product? Who will be a good hire from these 50 candidates? Who will get diabetes in the next 20 years? Who is trying to ruin my online reputation? Who is trying to commit a computer fraud? Who will vote for me? and a million others, then you can make a lot of money. To make these predictions though we need Deep data, Big Data isn’t good enough.
So how difficult is it to predict human behavior? Sometimes easy, sometimes hard. When Deep Mind’s Go-playing AI beat the world’s best oriental players it did so in the context of an environment that was extremely controlled and bound by a very tight set of rules. The behavior of the human player was limited to what was defined by the rules of the game and the computer could predict what was the best move for it and the best move for its opponent by being able to evaluate more moves faster than the human. With that restriction AI can win. The human could of course have “won” by just unplugging the computer! Deep Mind’s Go playing computer program is a machine. Machines are specialists. They do one thing very, very well. Your computer as a machine has a much better memory than you do in many ways. Indeed, there are many things your computer can do better than you. But so can a hammer or a screwdriver! Try drilling a screw into wood with your hand. Hammers and screwdrivers are also machines. So are computers. So are applications of machine learning algorithms.
AI is not a tool, its not even a tool set, its a tool factory. Within it you can make very good machines – hammers and screwdrivers, adequate for many different tasks. Many of these tools we call predictive models. Unfortunately, these predictive models don’t make themselves. People have to make them. And just like making real tools, making a good predictive model is a skill, where some people are better than others: some people have more experience, some deep domain specific intuition about the thing they’re modeling but low technical expertise, others very good computer skills but not much ability to understand business etc. The concept of data scientist is the idea that a special breed of people can bring together the distinct skills needed to build good predictive models and can do so by taking a data science degree. Becoming a good modeler however, takes time, experience, training and insight.
Data science, or data driven science, can provide an innovative framework in which business problems may be addressed, and AI can provide many useful tools. There are many modeling platforms with a plethora of such AI tools – SAS, SPSS, SAP, PowerAI – to name just a few. They can provide hammers and nails but they don’t build furniture out of them. Just like driving a Ferrari, if your grandmother is driving it you’re less likely to get to your destination, and certainly less quickly, than if Michael Schumacher is in the driver’s seat. Such platforms don’t provide solutions. They provide foundations upon which solutions can be built – if you have the right human resources to build them.
We need data scientists and AI/Machine Learning to predict human behavior in business because, at the heart of business, in fact, all human activity, is the concept of a decision: Customer decisions, business decisions, government decisions etc. However, for every decision we make there is an associated prediction behind it. If you decide to buy a particular hardware infrastructure its because you predict that it will lead to benefits for the business. If you design a particular digital marketing campaign it’s because you predict that it will lead to more customers.
Of course, human beings are already pretty good predictors of the behavior of other human beings. Evolution has designed us to be good at it – predicting when someone is angry or scared for instance. A good salesman can also predict who might be a good sales opportunity, using body language indicators, psychology and data about the customer. A good CIO might predict correctly what IT infrastructure would lead to a competitive advantage over their competitiors. All of these decisions however, come from using heuristics – hunches, guesses – based on experience and using limited data sets, often only what we see (e.g., read) or hear (e.g., what others tell us). We’ve been designed by evolution to understand such signals. The semantics of language is one of the most important. However, none of us speak or read “database”.
We are generating up to a zetabyte (1023 bytes) of data worldwide every year, and the vast majority of it is in “languages” we don’t speak or understand, and I don’t mean human languages. When you read a memo from the company CEO you understand it very well. You know what it is telling you. Maybe your budget was increased by 15%. When you look inside your company’s databases though – What is the data telling you? We can measure how much data is in there – the number of bits – we can even measure the amount of information in it, using information theory – we have no way though of measuring how much actionable knowledge is in it, because we don’t speak “binary”. We don’t know a priori if it’s got lots of useful information or hardly any. Creating a predictive model to answer a specific question of interest to the business is a way to test that. If the model can predict who will buy your product and why then that’s useful knowledge. Just like the CEO can ask you a multitude of things that could then potentially improve the business, so then with your databases there are a multitude of things that could be asked and tested and used to improve the business.
Data semantics is by far the most difficult task we face when dealing with the data revolution. We have concentrated on the technology of data processing, data storage, data accessibility and data communication and made huge progress. This is the key though – all of these advances consider data as data, as bits. None of it considers which data is important or not. Whether you use a data warehouse or a data lake is irrelevant from the point of data semantics. Your data has no value without being able to turn it into knowledge and, subsequently, predictions and decisions. So why have we made so much technological progress and not much on data semantics. Because data “plumbing” is easy – store data at point A, send it to point B etc. Data mining in that sense is a useful term because it hints at the fact that not all data is equally useful. However, to determine which data is useful, besides having a definition of useful, we also need to understand what the data means – data semantics. Slowly but surely future scientific and technological developments will be driven more and more towards understanding the meaning of data. Already important progress has been made using AI in, for example, language translation or text mining. IBM’s Watson system is a well known example. These systems have no understanding of the true meaning of the data they are considering though. What is needed is better and more efficient interfaces between human and machine to take exploit the advantages of both at the same time.
So, what does this all mean for the CIO? As the senior executive responsible for the IT and computer systems that support enterprise goals, it is your responsibility to use it, in tandem with other business areas, to produce tangible benefits for the business, such as higher ROI, more sales, more efficient operations etc. Remember, though, hardware and software don’t do anything on their own. The fuel that powers them is data. Data, in and of itself, doesn’t produce tangible benefits either. The data has to be turned into information, which then has to be turned into knowledge, which has to be turned into actions, i.e., decisions, i.e., predictions, which benefit the business. The data is useless unless it is incorporated into the decision making process. The question is then: How do you incorporate it into that process? It may be that the CMO wants to know the impact of last quarter’s digital marketing campaign. An analyst might then generate a report from some canned software to show how it went along different channels and show on another graph sales in the same period. This is descriptive analytics, business intelligence. You might be more sophisticated and have in place a predictive model for determining the profile of those prospects most likely to buy your product. You can then examine how well the model predicted the subsequent sales after the marketing campaign. This is predictive analytics. However, you could also have a model that predicted the Who, What, Where, When and Why for the campaign. Who to target in the campaign; What content to put in it; Where to target it – which websites for example; and When. This model is proscriptive, telling you exactly what actions/decisions to take based on predictions about the future outcomes of those decisions. This is where the future lies and it is an incredibly exciting one.