Consider the Machine-Learning Approach as You Dive Into AI

In a recent Accenture report, 75% of C-suite executives believe if they don’t scale artificial intelligence (AI) today, they’ll go out of business in five years. However, confusion or wrong decisions can result from jumping in without a good understanding of what AI can and cannot do, and how best to apply AI for the challenges in the field of advertising.

Researchers at Syracuse University have been collaborating to apply machine-learning techniques to analyze Comscore TV Essentials data with the goal of predicting linear television audience behavior. Dr. Chilukuri K. Mohan, professor of electrical engineering and computer science at Syracuse University, and author of Elements of Artificial Neural Networks, has been a lead researcher on this project and has some advice for marketers who are keen on entering this space.

Professor Mohan identified some challenges marketers face when embarking on ML projects. First, the true nature of the problem being addressed must be defined: classification, grouping, prediction, data mining, pattern recognition, optimization or anomaly detection. ML may not be necessary for questions that can be answered easily through other analytic approaches. If you are asking questions such as, “Why did an event happen?” you should use tools that aid in diagnostics or descriptive analyses that do not require ML. On the other hand, when you are looking for prescriptive or predictive answers, which must be constructed on the basis of existing data for which answers are known, a machine-learning approach would be more appropriate. (Also, ML models tend to be highly nonlinear; hence, if the problem is simple, statistical methods may be able to give you satisfactory answers.)

Understand the nature of data

Too often we have a dataset and simply plug it into our ML tools. But define how the data should be interpreted by the machine: How many attributes are numeric and/or nominal, as well as how many data points are available. For many problems, labeled data are essential, to enable training the ML model as well as evaluating the results of training. 

The ML model (a neural network) and its “hyperparameters” (the number of hidden layers and nodes in a neural network) must be carefully selected, preferring parsimonious models in order to reduce the risk of overfitting (on training data) and poor generalizations. For instance, a neural network with a hundred neurons (computing nodes) may perform better on new data than a neural network with a thousand neurons; even the latter appears to be best on available labeled data. The ML algorithm used to train the model must then be chosen, along with its own hyperparameters, addressing trade-offs between computational effort and the quality of results obtained. Best results are often obtained using an ensemble of models, such as basing the decision on 10 separately trained neural networks rather than a single one. This is analogous to consulting multiple experts who provide advice independently before making a final decision.

Machine learning offers marketers tremendous promise in making better decisions, given the copious amounts of data available and the rapid improvements in AI technology over the past few years. But it is not a panacea to the challenges faced by marketers. When it comes to advertising, the advertiser is solving people problems and the data scientist is solving math problems. The two need to come together at the outset and understand what decision the project is trying to result in and what tools are available to make those predictions. As with everything, a true understanding of each perspective will lead to better decisions and outcomes.