By John Mruz, VP of Innovation
In 1996, Bill Gates famously said that “If your business is not on the internet, then your business will be out of business.” More recently, it was Mark Cuban who said, “the world’s first trillionaires are going to come from somebody who masters AI and all its derivatives and applies it in ways we never thought of.” But with all of the media hype around artificial intelligence, what does it mean, specifically, for a business to be “using AI”?
October 2016 will likely be recognized as the point in history when AI went mainstream. First, President Obama released “The Administration’s Report on the Future of Artificial Intelligence”, a survey on the current state of AI, its potential applications, and impact on society. That was followed by a 60 Minutes segment featuring correspondent Charlie Rose conversing with a robot named Sophia (complete with facial features that were eerily humanlike) about her two goals: becoming smarter than humans and immortal. Since that time, AI has been front page news across the mainstream media landscape, including feature articles in the NY Times, Washington Post, Newsweek, and Fortune.
The unwanted byproducts of increased media attention (“the hype cycle”) are ambiguity and urgency. Ambiguity is a paradoxical relationship: the more attention AI receives, the less it is actually understood. Urgency is the corporate form of peer pressure: everybody else is doing AI so we need to be doing AI. When you combine ambiguity and urgency, the result is a marketing land grab: a lot of companies claiming to be doing AI – and perhaps truly believing they are doing AI – but, in reality, not doing AI at all. Thus is born the latest in a long line of hollow Silicon Valley buzzwords that trigger our collective BS meters.
Against this backdrop, allow me to talk very specifically about Rocket Fuel’s AI and how we use it. Artificial intelligence was THE foundational building block of Rocket Fuel’s entire business model. Rocket Fuel was founded in March 2008 with a vision of transforming the digital advertising industry through big data and artificial intelligence. Our founders recognized that the massive quantities of impression data we observed in fixed bidding on ad networks (and, later, the explosion of real-time bidding), combined with the ever-decreasing cost of computing and data storage, presented an opportunity to introduce intelligent computer algorithms to optimize media buying performance. In short order, our algorithms were able to outperform humans doing the same task manually, and in a fraction of the time.
But what exactly does Rocket Fuel mean by “artificial intelligence” and how does it work? Rocket Fuel subscribes to the generally accepted definition of AI presented by John McCarthy, one of the pioneers of the field: “AI is the science and engineering of making intelligent machines, especially intelligent computer programs.” In this context, “intelligence” refers to the computations needed to perform something that previously could only be done by humans. In order to build intelligent machines, you need three things: (1) a real-world objective; (2) a computational construct – or “model” – that generates the desired behavior; and, (3) a set of real-world data with which to train and validate the model.
It is important to recognize that there is no one-size-fits-all AI solution. The multitude and diversity of objectives to which AI can be applied (e.g. self-driving cars, adaptive thermostats, personal assistants, image recognition, text-to-voice, chat bots, robots) requires a variety of computational models. In this way, the world of AI can be viewed as a multi-limb tree, with each branch being a particular modeling technique (e.g. decision trees, linear/logistic regressions, neural networks, clustering, collaborative filtering, natural language processing, path planning, image recognition) to accomplish a particular objective.
Let’s dig deeper into Rocket Fuel’s objective, modeling, and data. Rocket Fuel’s real-world objective is to make marketing meaningful by predicting the potential of every moment so that marketers can deliver amazing experiences that get real results. By “moment”, we refer to the exact instant in time when a brand is presented with an opportunity to connect with a consumer. In our world, that moment can be when a visitor arrives at a web site owned by the brand or on a 3rd party publisher’s site that allows brands to advertise on it. When a moment presents itself, Rocket Fuel usually has less than 100 milliseconds to leverage all available data about the visitor to make the best decision and execute the ideal delivery of the marketer’s message. Success is achieved when we delight the visitors into taking an action.
One of the AI modeling techniques we use to solve this problem is called logistic regression. Logistic regression is a classic AI modeling tool that is best used for “classification” problems – for example, distinguishing a spam email from a regular email or a malignant tumor from a benign tumor. At Rocket Fuel, we use logistic regression to classify a marketing moment as a likely “conversion event” (i.e. will the person respond favorably to the ad) or not. The models use a wide array of data, including patterns of user behavior across the web, first party online (i.e., from the brand’s web site) and offline (i.e., from the brand’s CRM systems) data made available with permission by the brands, and segment data from 3rd party providers.
To give you some better insight into how we use logistic regression with this data, consider the following visualization. The X’s represent historical marketing moments where conversions occurred; the circles are marketing moments where no action was taken by the consumer. The model’s job is to establish a decision boundary that clearly demarcates the converters from the non-converters with minimal error. Behind this line is an equation that weights all of the features – that is, the data descriptors – that we know about each user and each moment; in this case, we present a simple boundary that is a function of only two features (e.g., Mobile Device Type and Age). Highly predictive features will have a greater impact on the placement of the decision boundary than less predictive features.
In the beginning of the campaign (left image), when there are fewer observations, the decision boundary can be very inaccurate. Here we see the decision boundary criss-crossing both populations, with many X’s and O’s misclassified. But as we observe more and more transactions, our model gets smarter and the accuracy of the decision boundary improves (center). In the ideal state, the model improves to the point where the decision boundary forms a clear demarcation that minimizes the risk of misclassifying a converter for a non-converter and vice versa.
The examples shown above are very simplistic – we show a simple decision boundary plotted against two features of interest for a single advertiser. But consider an example where you have millions of observations, the decision boundary is affected by thousands of features, and you are trying to optimize the potential of every marketing moment across thousands of campaigns. That is the scenario Rocket Fuel faces over 200 billion times per day. This is the type of problem that is tailor-made for artificial intelligence.
Hopefully this blog has cut through some of the “AI hype” and provided you with a better intuitive understanding of how Rocket Fuel uses AI to predict the potential of every moment. In future blog posts we will discuss the evolution of AI and the exciting potential to share intelligence between AI systems.