The AI technique may soon predict consumer responses, but don’t ditch all of your traditional marketing tools just yet

By Glen Urban, Artem Timoshenko, Paramveer Dhillon, and John Hauser

Improving market share, sales, and return on marketing investments is the trifecta of marketing organizations — and it’s just as elusive as winning the lottery. Some businesses are now asking whether deep learning (DL) models based on AI can help achieve marketing goals.

It may seem straightforward: If businesses can predict sales and consumer response to a variety of marketing strategies they are more likely to find a successful campaign model. But the reality is more complex.

Traditionally, consumer response was determined by linear models relating sales, or outcome variables, to price, promotion, advertising, and distribution measures known as explanatory variables.

Explanatory variables include a full range of marketing measures such as exposure to ads on the Internet, clicks on web sites, search efforts, and recommendation ratings, as well as traditional demographics, loyalty, price, advertising spending, promotion, distribution extent, and product attributes. Statistics determine whether the results are “significant,” meaning they are unlikely to have occurred by chance.

Fresh Approach

Deep learning takes a completely fresh approach to determining consumer response, in three key ways:

1 . First, it does not rely on a single, easy-to-interpret equation, but rather on a series of linear and non-linear transformations each representing a “neural” layer that is linked to the next layer. When there are many intermediate (hidden) layers, the method is called deep learning. DL models often require large data bases to estimate the many parameters.

2. Secondly, DL methods are judged by their ability to make accurate predictions from a new data base not just to fit an existing model. A DL model is good if it predicts well.

3. The third difference is structural advantage. DL can handle high-dimensional data that classical methods cannot — such as large numbers of marketing actions, text, images, audio, and video. Through its many layers, the DL model translates high-dimensional data into a form that can be processed to predict choices. DL can handle inputs such as comments, images, and credit card promotions, and the inputs need not be standard across all stimuli, as in market research. DL models also can merge large databases, such as raw clickstream data, with smaller databases such as consumer product ratings. We call such heterogeneous and highly dimensional data, “rich data.”

DL can also learn across applications. For example, we might train a DL model on a very large data base (such as Google News) and then modify the model for a particular application, say identifying customer needs.

The completeness allowed by heterogeneous data inputs and modeling response through multiple hidden layers argues for serious consideration of DL in marketing analytics.

Caveats and Prospects

Despite these strong attractions, however, there are caveats to consider before switching to DL for many marketing analytic applications.

While DL improves prediction relative to traditional models for classical marketing analytics data, the improvement is small and, perhaps, not yet sufficient to overcome DL’s disadvantages. One major disadvantage, for instance, is the inability to easily tell which variables drive the biggest response because the known variables are translated through many latent layers to predict choice. Additionally, DL models require substantial computing power and long run times.

Although DL currently falls short of offering a compelling case as a full substitute for classical response analytics, the prospects are exciting. DL has structural advantages over regression in analysis of the rich data bases that are now emerging in marketing.

Rich contextual data, such as verbal comments and visual images, are widely available user-generated content (Amazon reviews, Instagram posts, Facebook posts and comments, company websites). When we applied DL to this type of online user-generated content we found that a mixed human-DL hybrid method identified customer needs as well or better than traditional interviews, but more quickly and at substantially lower cost. [1]

Recently, one of our DL models was effective at predicting the aesthetic appeal of automobile images and could generate new, high-appeal external car designs to seed the design team’s creativity.[2] Another DL model was effective at modeling the dynamic evolution of customers’ news consumption preferences while also being predictive of customer return visits.[3]

Finally, methods of monitoring brand images from pictures and consumers posts are being used to predict which apparel items have high return rates, and identifying which coupons will lead to the highest incremental profit.[4] New applications of DL are emerging rapidly.

We are optimistic about DL’s ability to improve state-of-the-art of marketing practices by enabling analysis of new rich data bases and integration of real-time experimentation. This use of DL in marketing analytics can improve estimates of market response so profit and marketing ROI can be maximized; reveal new, creative opportunities for copy, media, and products, and, allow analysts to tackle new problems in product design, distribution, and media optimization.

Glen Urban is the David Austin Professor in Marketing, Emeritus at MIT Sloan School

Artem Timoshenko is a Ph.D. Candidate in Marketing at MIT Sloan

Paramveer Dhillon is a Postdoctoral Associate in Management Science at MIT Sloan

John Hauser is the Kirin Professor of Marketing at MIT Sloan

[1] Timoshenko, Artem and John R. Hauser (2018), “Identifying Customer Needs from User-Generated Content,” Marketing Science, 38(1):1–20.

[2] Burnap, Alex, John R. Hauser, Artem Timoshenko, “Generating and Testing Product Aesthetics: A Human-Machine Hybrid Approach,” MIT Management 2019.

[3] Dhillon, Paramveer and Aral, Sinan, “Modeling Dynamic User Interests: A Neural Network Approach,” MIT Management 2019.

[4] For example, Liu, Liu, Daria Dzyabura, Natalie Mizik, “Visual Listening In: Extracting Brand Image Portrayed on Social Media,” New Economic School.

Watch for a detailed article about this research in an upcoming MIT Sloan Management Review article.