Is Deep Learning a Game Changer for Marketing Analytics?

Companies are already making sophisticated marketing decisions with data and analytics. Will deep learning enable a leap forward — or just marginal gains?

Deep learning is delivering impressive results in AI applications. Apple’s Siri, for example, translates the human voice into computer commands that allow iPhone owners to get answers to questions, send messages, and navigate their way to and from obscure locations. Automated driving enables people today to go hands-free on expressways, and it will eventually do the same on city streets. In biology, researchers are creating new molecules for DNA-based pharmaceuticals.

Given all this activity with deep learning, many wonder how the underlying methods will alter the future of marketing. To what extent will they help companies design profitable new products and services to meet the needs of customers?

The technology that underpins deep learning is becoming increasingly capable of analyzing big databases for patterns and insights. It isn’t difficult to imagine a day when companies will be able to integrate a wide array of databases to discern what consumers want with greater sophistication and analytic power and then leverage that information for market advantage. For example, it may not be long before consumers, identified via facial recognition technology while grocery shopping, receive individualized coupons based on their previous purchase behavior. In the future, advertisements may be individually designed to appeal to consumers with different personalities and be delivered in real time as they view YouTube. Deep learning might also be used to design products to meet consumers’ personal needs, which could then be produced and delivered through automated 3D printing systems.

Different types of organizations will try to harness the powers of deep learning in their own ways. An automaker might use them to target new customers, revamp the buying process, or fine-tune product features a specific set of buyers will want. It could draw on a sea of relevant data to do all this, including auto repair data, consumer ratings on vehicle quality and reliability, car registrations, Twitter posts concerning the car-buying and user experience, Facebook posts showing people with their cars, manufacturers’ consumer relationship management data files, and clicks on the internet. A bank, meanwhile, could leverage deep learning to develop new products or services and customize its promotions. By analyzing data on customer loan histories, credit card transactions, savings and checking account records, website clicks, social media behavior, product ratings, and search histories, it could gain insights into the things certain customers value. What do 40-year-old professionals living in urban neighborhoods want most in a credit card? Do they prefer travel rewards, buyer protection, cash back, or low interest rates?

To be sure, a lot of managers already use analytics with statistical models and focused databases to track brand performance, schedule promotions, and make spending decisions. So how is deep learning different? Is it a fundamental leap forward, or will it simply enable marginal gains? In this article, we will examine these questions in relation to a study we conducted involving credit cards. In addition, we will consider what this research suggests about the future direction of analytics.

Although we are still in the early days of deep learning, it isn’t too early to ask: What will it offer companies compared with the existing analytics methods managers are accustomed to? Can it provide better predictions, and if not, how can it be improved? And what kinds of investments in data and technology will companies have to make in deep learning to take advantage of the latest and most powerful capabilities? Our research suggests that, while deep learning may not lead to large gains in predictive accuracy right away or in every setting, there are reasons for optimism.

Our Experiment

To compare deep learning with traditional methods for marketing analytics, we studied a large database of click-streams, demographics, and ad exposures relating to the credit card market from NerdWallet, a large online vendor of credit cards, based in San Francisco. We wanted to see if a multilevel deep learning model could predict credit card choices more accurately than traditional models.

The ability to predict customer choice is the first step toward improving decisions that go into product design, media resource allocation, how to promote the product (in this case, a credit card), and whom to target. Knowing what people value most requires experimentation and predictive choice modeling. Our hunch was that deep learning would provide a clearer, more useful picture than a simpler regression model. To test that assumption, we looked at the credit card selection processes of 260,000 individuals, taking into account 25 demographic factors (including obvious things like age, gender, and household income, and more detailed categories such as credit score, cards the consumer currently owns, and ZIP code); 132 attributes for each card (such as interest rate, whether the card offered reward points, travel miles, or cash back, and card fees for annual membership and balance transfers); and the cards each person applied for.

The NerdWallet website aggregates information and expert reviews about some 2,000 different credit cards from hundreds of banks so that customers can compare features and decide what’s important to them. The products that users view during their online sessions are recorded. NerdWallet also notes when a person clicks to see more details or selects particular cards for comparison. After users decide which credit card they want from the NerdWallet site, they are directed to the corresponding bank’s website to complete an application. For the purposes of our study, we treated this action as an indication of eventual choice.

Using NerdWallet’s database, we compared three choice models. (For more details, see “Technical Note for Analysts.”) The first model was a straightforward linear regression of choice as a function of the individual user demographics and card attributes. Each variable had a simple direct effect on choice through one equation, with no interactions among variables. For example, a coefficient of .058 times the number of reward miles granted per thousand dollars charged by the individual might help predict whether a consumer would choose the American Express Green Card.

The second model was a simple deep learning model, which incorporates hidden layers between the input variables and the probability of choice. Each input variable (for example, reward miles) is linked to a latent variable, which in turn is tied to the probability of choice. Latent variables aren’t specifically identified; rather, they are combinations of input variables at the first level. They link to other latent variables in the next model level, and subsequent latent variables are combinations of other latent variables. This may seem unnecessarily complex, but linking latent variables across hidden layers is what gives deep learning its power (and even its name). Our model had three levels. The first layer captured the effects of the observed card attributes and demographics on a set of latent variables, and then we looked at the impact of those latent variables on subsequent layers of latent variables. The hope was that with three hidden layers that allowed for complex interactions and nonlinearities, we could predict card choice more accurately.

Our third model was an enhancement of the simple deep learning model, where the dependent measures were the probability of choice and the particular cards consumers considered in their final choice. It took into account granular information about the buying process and the other credit cards people considered (based on their click-streams). By adding this additional step of consideration to the buying process, we thought we might improve on the accuracy of the choice predictions of the basic deep learning model.

What We Found

Based on our analysis, the two models that drew on deep learning were able to predict credit card choices with more precision than the traditional approach. (See “How Deep Learning Can Outperform Traditional Marketing Analytics.”) But the improvements were not as large as we thought they would be.

The simple regression model had an accuracy rate of 70.5%, meaning that in 70.5% of the cases, we were able to predict correctly which card a particular consumer would apply for and which ones he or she would not apply for. The simple deep learning model was slightly more accurate, at 71.7%, and the more sophisticated model had an accuracy rating of 73%.

We had thought that adding more latent layers would result in more dramatic improvements. But our expectations may have been overblown. Well-designed linear regression models refined by experience can be robust and accurate. The many parameters — that is, the many demographic and card attribute measures — enabled good predictions from the traditional approach.

Costs and Benefits

So what did our study tell us about the potential for deep learning? Does it make sense for marketing organizations to invest in the technology and develop the critical capabilities now, given how small the improvements in choice prediction may be? In our opinion, the small gains in predictive accuracy over what’s possible using traditional data are not likely to generate sufficiently high returns in most cases to justify significant investments. Before simply diving in, organizations need to look beyond the potential gains and consider the challenges and costs of implementation.

A major drawback of deep learning is that it’s difficult to identify which variables drive the biggest response. Was it the credit card travel reward levels or the low interest rates that had a bigger influence on consumer choice? Because the variables are processed through so many latent layers affecting choice, it’s hard to measure the impact of one factor. We can simulate the effect of changes in one variable, but the effect depends on the levels of all the other variables used in the complex structure of the model. It can be difficult to link intuitive judgments of marketing managers to the model results, and this will make implementation difficult.

There are other factors to take into account. There are costs to acquire deep learning technology, staff required to implement it, and additional data costs that may need to be incurred. Finally, even with today’s fast computers, deep learning models need huge amounts of computer power and may experience long run times. This becomes a real limitation if real-time implementation is required. The reality is that in most situations a compelling case for deep learning as an alternative to traditional data analysis can’t be made unless the modest gains in prediction can generate big cost reductions or profit opportunities.

That said, we still think the prospects for deep learning are bright in some contexts. As we have noted, it has significant advantages over traditional regression for analyzing “rich” databases that include images and nonnumerical data. Rich data could include user-generated content (such as Amazon reviews, Instagram posts, Facebook posts, and comments on company websites), and the potential value of such data could be significant. Although the NerdWallet site contained highly granular search and product data, it didn’t rise to the level of rich data, so the advantages didn’t come into play. Having consumers’ verbal reviews of credit cards may have improved the predictive accuracy.

Deep learning is particularly well suited to identifying patterns in heterogeneous and unstructured data. Recently, for example, some of us developed a deep-learning-based model to help automotive designers predict how consumers would react to prototypes on the drawing board while also generating ideas for new products.1 In another case, researchers created a model from customers’ news-consumption behavior to predict future viewing habits.2 Lately, companies have begun to use deep learning to analyze pictures of branded apparel that consumers post on social media and anticipate which items will have the highest customer return rates and which coupons will generate the most profit.3 This allows managers to develop better product-return policies and more effective discount strategies. A prior study by some of us showed that deep learning methods helped companies prescreen user-generated content more efficiently, thereby focusing analysts’ efforts on highly informative reviews that identified customer needs at substantially lower costs than the traditional methods.4

New opportunities for deep learning applications are emerging all the time. Deep learning encourages experimentation and enables real-time adaptability of A/B experiments and tweaking as results are observed. We already have developed a simple AI approach that adapts the images and content of online ads to match a consumer’s cognitive style and designs real-time effective experimentation systems to adaptively learn the best ad copy for each group of customers. Existing statistical models let companies target ads to online consumers. But the AI system allows marketers to tailor ad copy to individuals’ cognitive and communication styles, customize pictures and graphics to suit their visual preferences, and incorporate stats and other evidence to align with their decision-making styles. The applications are showing promising results and are likely to deliver even better results as new deep-learning-based algorithms and more comprehensive databases become available.5

Implications for Companies

What will all this mean for companies? Based on our experience, we see several takeaways for marketing managers:

1. Be alert to the future opportunities. Established statistics-based systems will still play a critical role in marketing analytics, but it’s not too early to imagine how deep learning can enhance current operations and enable solutions to new problems. High potential areas include more sophisticated promotion budgeting and scheduling, more advanced targeting of customers, and more refined new product development.

2. Create rich databases. Integrating many sources and types of data will likely produce major advances in response analysis and the optimization of marketing resources. The biggest costs of deep learning are apt to involve data acquisition and database creation and integrity. In our study of credit card selection, more than 50% of the costs were associated with data collection, labeling, and cleaning. So managers who deem the benefits of deep learning worth exploring in their business should invest in software to gather consumer comments from their website, Amazon, Twitter, and Instagram along with individual click-stream records of search and purchase decisions by customers.

3. Build capability through training and new hires. To take advantage of opportunities in deep learning, you will need in-house experts capable of working with your databases. They will need to know how to maximize the use of rich data, employ it to solve previously unsolved marketing problems, and find new strategic insights. Processing rich data requires advanced knowledge in AI and budgets to hire consultants to augment existing internal staff.

4. Develop plans for experimentation. Beyond using deep learning to conduct better and more efficient A/B experiments, you can also use it to test new marketing variables. The greatest potential may lie in focused experiments that systematically employ different variables in order to measure the response. The state of the art would be to apply different real-time stimuli for each user. The experiments could then be analyzed via deep learning methods.

Thoughts About the Future

Despite the costs and challenges of implementation, we are optimistic about deep learning’s potential to advance state-of-the-art marketing practice by enabling analysis of rich databases and real-time experimentation. This application in marketing analytics can: (1) improve estimates of market response to help maximize profit and marketing ROI, (2) reveal new opportunities for product development, and (3) allow for more targeted product design, distribution, promotion, and media optimization. Although only marginal gains are available with old data and traditional analytics, in many new analytic applications, the improvements in accuracy and insight will justify the investment in deep learning technology and data.

References

1. A. Burnap, J.R. Hauser, and A. Timoshenko, “Design and Evaluation of Product Aesthetics: A Human-Machine Hybrid Approach,” working paper, MIT Sloan School of Management, Cambridge, Massachusetts, July 2019.

2. P. Dhillon and S. Aral, “Modeling Dynamic User Interests: A Neural Network Approach,” working paper, MIT Sloan School of Management, Cambridge, Massachusetts, 2019.

3. L. Liu, D. Dzyabura, and N. Mizik, “Visual Listening In: Extracting Brand Image Portrayed on Social Media,” working paper, Social Science Research Network, 2019.

4. A. Timoshenko and J.R. Hauser, “Identifying Customer Needs From User-Generated Content,” Marketing Science 38, no. 1 (January-February 2019): 1-20.

5. G. Urban, J. Hauser, G. Liberali, et al., “Morph the Web to Build Empathy, Trust, and Sales,” MIT Sloan Management Review 50, no. 4 (summer 2009): 53-61.

i. For further information, see I. Goodfellow, Y. Bengio, and A. Courville, “Deep Learning” (Cambridge, Massachusetts: MIT Press, 2015); A. Géron, “Hands-On Machine Learning With Scikit-Learn and TensorFlow” (Sebastopol, California: O’Reilly Media, 2017); and R.S. Sutton and A.G. Barto, “Reinforcement Learning: An Introduction” (Cambridge, Massachusetts: MIT Press, 2018).

Acknowledgments

We would like to thank Suruga Bank and MIT’s Initiative on the Digital Economy for their funding support, and NerdWallet and Comscore for providing the data used in the research.